Voice-controlled super-resolution ultrasound imaging and reporting powered by multimodal large language models Authors: Ning Guo, Zixuan Deng, Qin Tan, Kai Sheng, Xuyang Wang, Shiyun Wang, Chen Hua Published: 2026 ABSTRACT Super-resolution ultrasound imaging (SRUI) surpasses the diffraction limit of conventional ultrasound, enabling visualization of microvascular architecture and hemodynamics with potential applications in neurology, oncology, and cardiology. However, clinical adoption remains limited by complex parameter optimization, subjective interpretation, and time-consuming workflows. We present a multimodal artificial intelligence framework that integrates a custom SRUI platform with large language models of DeepSeek-R1 for natural language processing and of MiniCPM-V for image recognition. Clinicians issue voice commands to initiate imaging tasks, which are translated into acquisition parameters, including temporal windows and adaptive microbubble filtration. The system performs super-resolution reconstruction, extracts quantitative vascular metrics, and generates structured diagnostic reports incorporating relevant clinical context. Filtration thresholds were dynamically determined using the Microbubble Similarity Score. Structured reports were generated within approximately four minutes. Evaluation by fourteen clinicians demonstrated good structural integrity and standardized terminology. This framework streamlines SRUI workflows and supports AI-assisted, clinically contextualized super-resolution ultrasound imaging. Trial registration: Chinese Clinical Trial Registry ChiCTR2100048361 registered July 6, 2021. Introduction Ultrasound imaging remains a cornerstone of real-time, non-invasive diagnostics across diverse clinical fields. Building upon this foundation, Super-Ultrasound for Greater Accuracy and Resolution (SRUI) has emerged as a promising imaging approach, enabling microvascular visualization beyond the conventional diffraction limit of conventional ultrasound1–4. This enhanced resolution ability is achieved through sub-wavelength tracking of microbubble contrast agents, allowing improved visualization of capillary-level hemodynamics3,5. Such capabilities position SRUI as a potential modality for neurological6, oncological7, and cardiovascular applications8. Despite this potential, SRUI’s clinical translation faces significant constraints. Wider adoption requires specialized operator expertise, sophisticated hardware-software integration, and time-intensive workflows for super-resolution data acquisition, processing, and interpretation—barriers to routine clinical accessibility and scalability9–11. Concurrently, artificial intelligence (AI) has increasingly reshaped clinical decision support, with AI-driven solutions demonstrating competitive performance in radiological and pathological tasks, including image segmentation, anomaly detection, and predictive analytics12,13. More recently, large language models (LLMs) such as GPT- and DeepSeek-series have expanded these applications, exhibiting proficiency in understanding and generating accurate clinical language to automate documentation, report drafting, and interactive interfaces14–16. Despite these advances, AI has yet to fully address the interconnected challenges specific to SRUI implementation. Current SRUI practice encounters ongoing implementation challenges. First, diagnostic-quality image acquisition demands manual parameter tuning and expert knowledge of scanning physics and signal processing. Second, image interpretation inherently requires correlating multimodal data with clinical context to reduce subjectivity, minimize inter-observer variability, and prevent diagnostic discrepancies. Third, report generation remains laborious and temporally disconnected from image acquisition, increasing clinician burden and potentially delaying care. Finally, SRUI accessibility is particularly limited in community and resource-constrained settings due to cost, complexity, and extensive training requirements. These collective challenges highlight the need for an intuitive, efficient, and standardized paradigm to leverage advanced ultrasound technologies. To bridge these critical gaps, we introduce a multimodal AI framework integrating our SRUI imaging platform with an LLM-powered conversational interface (Fig. 1). This integrated system enables clinicians to initiate and control super-resolution studies using natural spoken commands such as “Measure hippocampal flow in the left hemisphere”. The LLM component dynamically parses these commands to configure acquisition parameters and coordinate microbubble localization. Processed contrast-enhanced ultrasound (CEUS) data subsequently generate super-resolution vascular maps, from which quantitative metrics—including velocity, microvascular density, and perfusion indices—are extracted, tabulated, and visualized. The LLM then synthesizes a comprehensive diagnostic report contextualizing imaging findings within the patient’s unique clinical history and prior examinations, providing structured interpretations and management considerations directly aligned with the clinician’s initial command.Fig. 1Integrated LLM-SRUI framework workflow.A Clinician issues voice command such as “Measure super-resolution hippocampal flow in the left hemisphere”. B, C Multimodal LLMs of DeepSeek-R1 and MiniCPM-V interpret command, configures acquisition parameters, and triggers SRUI platform. D Microbubble localization with dynamic filtration thresholds, with super-resolution microvascular map generation and quantitative feature extraction of velocity, density and tortuosity. E AI-synthesized diagnostic report with clinical interpretations and management recommendations. CEUS Contrast-enhanced ultrasonography. Integrated LLM-SRUI framework workflow. A Clinician issues voice command such as “Measure super-resolution hippocampal flow in the left hemisphere”. B, C Multimodal LLMs of DeepSeek-R1 and MiniCPM-V interpret command, configures acquisition parameters, and triggers SRUI platform. D Microbubble localization with dynamic filtration thresholds, with super-resolution microvascular map generation and quantitative feature extraction of velocity, density and tortuosity. E AI-synthesized diagnostic report with clinical interpretations and management recommendations. CEUS Contrast-enhanced ultrasonography. Results Intelligent parameter determination via natural language processing The integration of our custom-made super-resolution ultrasound imaging (SRUI) platform with large language models, DeepSeek-R1 for text processing and MiniCPM-V for image processing, demonstrated the ability to autonomously determine critical super-resolution imaging parameters based on clinician voice commands (Fig. 2). This section details the system’s performance in translating natural language instructions into the appropriate imaging parameters required for generating diagnostically interpretable SRUI images and reports. Two core parameter categories were successfully managed. The first one is elementary parameters input, such as temporal selection windows, while the second one includes more advanced parameters input, like dynamic microbubble filtration thresholds.Fig. 2Integrated LLM-driven architecture for automated super-resolution reconstruction and structured clinical report generation.The end-to-end system architecture begins with clinician voice commands, which are converted to text and processed through natural language understanding by locally deployed large language models including DeepSeek-R1 and MiniCPM. The NLP module first validates command feasibility and generates structured reconstruction parameters. These parameters dynamically configure the MATLAB-based super-resolution pipeline, including frame-wise deconvolution, motion correction, adaptive band-pass filtering, microbubble localization, and tracking across the entire contrast-enhanced ultrasound dataset. Key quantitative microvascular parameters extracted from the reconstruction process are exported as structured data files in CSV and XLSX formats. Simultaneously, the reporting module generates standardized clinical reports based on HTML or LaTeX source templates, which can be exported in DOCX, PDF, or web-based HTML formats for clinical documentation and mobile viewing. The complete pipeline integrates parameter inference, microbubble reconstruction, quantitative analysis, and automated reporting into a unified workflow, achieving fully automated voice-to-report super-resolution imaging within approximately four minutes for a 5000-frame CEUS acquisition, without manual parameter tuning. Integrated LLM-driven architecture for automated super-resolution reconstruction and structured clinical report generation. The end-to-end system architecture begins with clinician voice commands, which are converted to text and processed through natural language understanding by locally deployed large language models including DeepSeek-R1 and MiniCPM. The NLP module first validates command feasibility and generates structured reconstruction parameters. These parameters dynamically configure the MATLAB-based super-resolution pipeline, including frame-wise deconvolution, motion correction, adaptive band-pass filtering, microbubble localization, and tracking across the entire contrast-enhanced ultrasound dataset. Key quantitative microvascular parameters extracted from the reconstruction process are exported as structured data files in CSV and XLSX formats. Simultaneously, the reporting module generates standardized clinical reports based on HTML or LaTeX source templates, which can be exported in DOCX, PDF, or web-based HTML formats for clinical documentation and mobile viewing. The complete pipeline integrates parameter inference, microbubble reconstruction, quantitative analysis, and automated reporting into a unified workflow, achieving fully automated voice-to-report super-resolution imaging within approximately four minutes for a 5000-frame CEUS acquisition, without manual parameter tuning. Clinician instructions specifying temporal ranges for contrast-enhanced ultrasound (CEUS) analysis were consistently and accurately interpreted. For commands such as “Show the super-resolution images from 90 seconds to 3 minutes post-injection,” DeepSeek-R1 reliably extracted the relevant time points and converted them into frame-acquisition parameters. The model achieved accurate translation of relative time descriptions, such as “early arterial phase”, “peak enhancement” and “late washout”, into absolute time ranges based on the known contrast injection time point. The system handled nuanced instructions involving discontinuous intervals (e.g., “Analyze the first 45 s and then from 2 min 30 s to 4 min”) by segmenting the request into discrete temporal blocks (t = 0–45 s; t = 150–240 s). This capability ensures the system focuses computational resources and analysis precisely on the clinically relevant phases of contrast kinetics as directed by the operator. Determining the optimal microbubble signal filtration threshold presented a more complex challenge due to inherent spatial and temporal heterogeneity in microbubble concentration within CEUS sequences. Our framework addressed this through a multi-stage analysis pipeline initiated by a pre-processing command like “Assess microbubble density and suggest filtration”. The LLM automatically wrote an embedded analytical module that processed the raw CEUS video. This module calculated microbubble density metrics for every frame, including spatial distribution maps, mean pixel intensity within regions-of-interest (ROIs), and temporal fluctuation profiles. This provided a quantitative baseline of contrast agent presence throughout the acquisition. Leveraging the pre-processed density data, DeepSeek-R1 dynamically adjusted the microbubble signal filtration threshold. For frames exhibiting very high microbubble concentrations, which is indicative of potential microbubble signal saturation or overlap, the LLM lowered the filtration threshold. This stricter filtering helped mitigate the “blooming” artifact and improved the discrimination of individual microbubble tracks, which is important for accurate super-resolution localization. Conversely, in frames with low microbubble density, a slightly higher or more lenient threshold was applied to retain sufficient signal points for meaningful super-resolution image reconstruction without introducing excessive background noise. To objectively evaluate the quality of the isolated microbubble signals used for super-resolution, the system employed a custom-defined Microbubble Similarity Score (MSS). This metric (scale: 0–1000) quantitatively assessed key characteristics of the detected microbubbles (size distribution, intensity profile, spatial clustering) against an ideal reference library derived from controlled phantom studies. A high MSS (e.g., 870/1000) indicated that the isolated signals closely matched the expected properties of intact, individual microbubbles, suggesting that the filtration parameters were effectively isolating true contrast agent signals while suppressing noise and artifacts. The MSS provided an iterative feedback loop, allowing the LLM to iteratively refine its threshold suggestions if the initial score fell below a pre-defined quality cutoff. For example, MSS < 750 triggered re-analysis. Based on the frame-specific concentration analysis and guided by the goal of maximizing the MSS, DeepSeek-R1 generated a recommended microbubble filtration threshold value (or a small range of values) tailored to the specific CEUS sequence. This recommendation was presented alongside the calculated MSS and a brief rationale. For instance, recommended threshold of MSS was set to be 870/1000 with optimal discrimination achieved and minimal background noise retained. The final output of DeepSeek-R1’s processing was the automatic generation of a structured parameter set (Table 1) required to execute the SRUI reconstruction algorithm. Detailed clarifications for each parameter listed in Table 1 are provided in the Document S1. This table consolidated both the temporally defined frame ranges and the optimized, sequence-specific microbubble filtration threshold. For instance, the command “Generate super-res image of the tumor vasculature from minute 1–2.5, ensure good microbubble separation” resulted in the following parameter set, with (1) Selected Time Range being 60 s–150 s post-injection, and (2) Microbubble Filtration Threshold being 0.31 (Dynamically optimized, MSS: 855/1000).Table 1Automatically configured SRUI super-resolution reconstruction parameters derived from LLM interpretation of clinician voice commandsItemValue_1Value_2Value_3Value_4ExplanationMotionCorrect_ynyMotion correction applied? (y/n)Frames Of Interests (FOI)18378959Frame range for motion correction and super-resolution; min and max FOI displayed; Frame of Interest (FOI), it is defined as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\rm{FOI}}=[{\rm{TOI}}\times {\rm{fps}}]$$\end{document}FOI=[TOI×fps] where TOI is the time of interest extracted by the DeepSeek-R1 model from the clinician’s natural language input, and fps is the frame rate of the contrast-enhanced ultrasound acquisition. This definition clarifies how the relevant time window is calculated based on the specified time of interest and the frame rate.Frame Rate FPS(Hz)27Acquisition frame rate (Hz)Time of Interests (TOI) in s68331.8Temporal window for motion correction and super-resolution; min and max TOI displayedDurationAll331.8518519Total video duration (s)ROI_MotionCorrectionEigen11.51190.51521.98558.98B-mode region selected for motion correctionContrastInLeftOrRightRContrast video position within frame: Left (L) or Right (R)max_shift100Maximum displacement threshold for motion correctionDistance_between_Bmode_Contrast550Spatial separation between B-mode and contrast imaging planesJermanVesselFilter_ynnJerman vessel filtering enabled? (y/n)SingularValueNumber1202100SVD decomposition diagonal element indices (upper bound ≤ frame count)neighbordistancez10Minimum inter-bubble distance in Z-dimension (height)neighbordistancex10Minimum inter-bubble distance in X-dimension (width)numberOfParticles50Estimated bubble count per framethresholdNoise0.9Frame-specific noise threshold valueimresize_factor1Bubble interpolation scaling factor (e.g., 1.65 for expansion)max_linking_distance6Maximum permissible linking distance for tracking (pixels)max_gap_closingInfMaximum allowable frame gap for track linkingzoomfactor20Image magnification factor (e.g., 20 × [910 1260])gaussian_size0.25Rayleigh disc/Airy speckle diametersmoothfactor0.45Vessel smoothness parameter (0-1, higher values suppress small vessels)colormax1.2Maximum intensity value for monochrome density colormapUpDownNormalRange-0.80.8Color range for bidirectional flow density mappingMaxFrameIntervalInfMaximum consecutive frames for unsmoothed velocity visualizationVelocityTypevabsVelocity component for unsmoothed plots: vabs, vx, vy, or thetaLineWidth1Line thickness for velocity trajectory plotsBackGroundColorblackBackground color for velocity vector plotsArrowOnOrOffoffArrow display toggle for velocity vectorsVelocityType_smoothvabsVelocity component for smoothed plots: vabs, vx, or vyVnormalRange_smooth-0.50.5Normalized velocity range for smoothed visualizationInterpolateTypesInterpolation method: s (spline) or L (linear)exportfilename/dssg/home/acct-***/medzyys4/EINSTEIN/BREEZE_revisitedV2023M0128/ResultsOutput directory path for resultsPure_video_nameWangFuGengSource ultrasound file identifierMotionCorrectedContrastMat_name./Motion_Correction/cropped_video/Corrected_Contrast_matrix_Wang***_20241020T153723.mat(Input) Path to motion-corrected contrast video (.mat)VesselTrack_name./tmp/VesselTrackCELL_***_20241020T155258.mat(Output) Path to vessel tracking data (pre-Kalman/interpolation)ROIsize522559ROI dimensions (width × height)Parameters include motion correction settings (e.g., frame ranges, spatial ROIs), microbubble processing thresholds (e.g., noise filtering, linking distances), and visualization specifications (e.g., velocity mapping, colormap ranges). Values were dynamically optimized using real-time microbubble concentration analysis and validated via Microbubble Similarity Score (MSS). Table 1Automatically configured SRUI super-resolution reconstruction parameters derived from LLM interpretation of clinician voice commandsItemValue_1Value_2Value_3Value_4ExplanationMotionCorrect_ynyMotion correction applied? (y/n)Frames Of Interests (FOI)18378959Frame range for motion correction and super-resolution; min and max FOI displayed; Frame of Interest (FOI), it is defined as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\rm{FOI}}=[{\rm{TOI}}\times {\rm{fps}}]$$\end{document}FOI=[TOI×fps] where TOI is the time of interest extracted by the DeepSeek-R1 model from the clinician’s natural language input, and fps is the frame rate of the contrast-enhanced ultrasound acquisition. This definition clarifies how the relevant time window is calculated based on the specified time of interest and the frame rate.Frame Rate FPS(Hz)27Acquisition frame rate (Hz)Time of Interests (TOI) in s68331.8Temporal window for motion correction and super-resolution; min and max TOI displayedDurationAll331.8518519Total video duration (s)ROI_MotionCorrectionEigen11.51190.51521.98558.98B-mode region selected for motion correctionContrastInLeftOrRightRContrast video position within frame: Left (L) or Right (R)max_shift100Maximum displacement threshold for motion correctionDistance_between_Bmode_Contrast550Spatial separation between B-mode and contrast imaging planesJermanVesselFilter_ynnJerman vessel filtering enabled? (y/n)SingularValueNumber1202100SVD decomposition diagonal element indices (upper bound ≤ frame count)neighbordistancez10Minimum inter-bubble distance in Z-dimension (height)neighbordistancex10Minimum inter-bubble distance in X-dimension (width)numberOfParticles50Estimated bubble count per framethresholdNoise0.9Frame-specific noise threshold valueimresize_factor1Bubble interpolation scaling factor (e.g., 1.65 for expansion)max_linking_distance6Maximum permissible linking distance for tracking (pixels)max_gap_closingInfMaximum allowable frame gap for track linkingzoomfactor20Image magnification factor (e.g., 20 × [910 1260])gaussian_size0.25Rayleigh disc/Airy speckle diametersmoothfactor0.45Vessel smoothness parameter (0-1, higher values suppress small vessels)colormax1.2Maximum intensity value for monochrome density colormapUpDownNormalRange-0.80.8Color range for bidirectional flow density mappingMaxFrameIntervalInfMaximum consecutive frames for unsmoothed velocity visualizationVelocityTypevabsVelocity component for unsmoothed plots: vabs, vx, vy, or thetaLineWidth1Line thickness for velocity trajectory plotsBackGroundColorblackBackground color for velocity vector plotsArrowOnOrOffoffArrow display toggle for velocity vectorsVelocityType_smoothvabsVelocity component for smoothed plots: vabs, vx, or vyVnormalRange_smooth-0.50.5Normalized velocity range for smoothed visualizationInterpolateTypesInterpolation method: s (spline) or L (linear)exportfilename/dssg/home/acct-***/medzyys4/EINSTEIN/BREEZE_revisitedV2023M0128/ResultsOutput directory path for resultsPure_video_nameWangFuGengSource ultrasound file identifierMotionCorrectedContrastMat_name./Motion_Correction/cropped_video/Corrected_Contrast_matrix_Wang***_20241020T153723.mat(Input) Path to motion-corrected contrast video (.mat)VesselTrack_name./tmp/VesselTrackCELL_***_20241020T155258.mat(Output) Path to vessel tracking data (pre-Kalman/interpolation)ROIsize522559ROI dimensions (width × height)Parameters include motion correction settings (e.g., frame ranges, spatial ROIs), microbubble processing thresholds (e.g., noise filtering, linking distances), and visualization specifications (e.g., velocity mapping, colormap ranges). Values were dynamically optimized using real-time microbubble concentration analysis and validated via Microbubble Similarity Score (MSS). Automatically configured SRUI super-resolution reconstruction parameters derived from LLM interpretation of clinician voice commands Parameters include motion correction settings (e.g., frame ranges, spatial ROIs), microbubble processing thresholds (e.g., noise filtering, linking distances), and visualization specifications (e.g., velocity mapping, colormap ranges). Values were dynamically optimized using real-time microbubble concentration analysis and validated via Microbubble Similarity Score (MSS). The LLM included internal consistency checks, flagging potential conflicts (e.g., a requested time range with insufficient microbubble signal for reliable super-resolution) and suggesting alternatives. This automated parameterization reduced the need for manual, trial-and-error adjustment by the sonographer, streamlining the workflow. Automatic generation of super-resolution scanning report: multi-modal data integration and image interpretation The SRUI algorithm processed raw ultrasound data from the Mindray Resona R9 system using the SP5-1U transducer, generating detailed microvascular maps of the hippocampus via the transtemporal window. Intravenous administration of SonoVue® microbubble contrast enabled visualization of vascular networks at sub-diffraction microvascular resolution. The system detected subtle pathological patterns, including elevated tortuosity (13.33 ± 2.15) and heterogeneous perfusion. DeepSeek, combined with MiniCPM-V exhibited the ability to correlate super-resolution images with clinical parameters. When processing the representative case shown in Fig. 3 (64-year-old male with uncontrolled diabetes), the model identified valuable phenomena for further diagnosis, including (1) increased vascular density in CA1-CA3 subfields (qualitatively described as “densely packed, tortuous microvessels”), (2) hemodynamic irregularities through velocity vector analysis (“heterogeneous flow velocities with preserved continuity”), and (3) morphological indicators of microvascular remodeling (“elevated tortuosity suggestive of vascular adaptation”).Fig. 3AI-Generated Diagnostic Report for Diabetic Microangiopathy.Representative structured report autonomously synthesized in ~30 s. A Patient data (A1) and scanner parameters (A2). B Quantitative SRUI metrics. C Clinical impression section contextualizing findings and management recommendations. AI-Generated Diagnostic Report for Diabetic Microangiopathy. Representative structured report autonomously synthesized in ~30 s. A Patient data (A1) and scanner parameters (A2). B Quantitative SRUI metrics. C Clinical impression section contextualizing findings and management recommendations. After that, DeepSeek synthesized imaging features, quantitative parameters, and clinical history into structured diagnostic reports. For the index patient, the system generated the key clinical correlation: “Moderately elevated vascular density (0.7257) and tortuosity (13.33) suggest compensatory microvascular remodeling in response to chronic metabolic stress. This pattern aligns with early diabetic microangiopathy given concomitant hyperglycemia (HbA1c 8.5%) and hyperinsulinemia.” The platform incorporated critical management recommendations including (1) glycemic optimization to mitigate microvascular progression, (2) multidisciplinary follow-up involving endocrinology and neurology, (3) interval super-resolution ultrasound reassessment at 6 months, and (4) correlation with neurocognitive testing The document generation module required approximately 20–30 s for report formatting alone, while the complete end-to-end reconstruction and report generation process required approximately four minutes for a 5,000-frame CEUS dataset. As demonstrated in the attached sample report (Fig. 3 and Document S2), MATLAB accurately executed a series of Word templating commands to generate a structured diagnostic document. The system integrated patient demographics and scanner parameters into predefined sections, embedded super-resolution images with standardized labeling for clarity and consistency, and formatted key quantitative metrics into well-structured diagnostic tables. In addition, institutional branding and regulatory compliance elements were automatically applied, and a detailed impression section was generated, providing recommendations aligned with imaging findings. Super-resolution visualization of intraplaque neovascularization To further evaluate the feasibility of the proposed framework in vascular pathology, super-resolution ultrasound imaging was performed in a patient with type 2 diabetes mellitus presenting with carotid plaque. As shown in Fig. 4A, SRUI enabled detailed morphological visualization of intraplaque neovessels, revealing a dense and spatially heterogeneous microvascular network within the plaque region. The super-resolved microvascular structures extended beyond the spatial resolution of conventional Doppler imaging and demonstrated irregular branching patterns consistent with neovascular proliferation.Fig. 4Super-resolution ultrasound imaging of neovascularization within carotid plaques in a patient with type 2 diabetes mellitus.A Morphological visualization of intraplaque neovessels. B Microvascular flow velocity mapping. Super-resolution ultrasound imaging of neovascularization within carotid plaques in a patient with type 2 diabetes mellitus. A Morphological visualization of intraplaque neovessels. B Microvascular flow velocity mapping. Fig. 4B presents the corresponding microvascular flow velocity map. Spatially heterogeneous flow velocities were observed within the neovessel network, with localized regions exhibiting relatively elevated flow signals. The combined structural and velocity information provided by SRUI allowed simultaneous assessment of microvascular architecture and hemodynamic characteristics within the plaque. These findings demonstrate the capability of the proposed LLM-SRUI workflow to acquire, reconstruct, and generate quantitative microvascular maps in carotid plaque, supporting comprehensive visualization of diabetes-associated intraplaque neovascularization within a streamlined imaging and reporting process. Comprehensive statistical validation of AI-generated super-resolution ultrasound reports Fourteen clinicians independently evaluated five AI-generated reports using a structured sixteen-item questionnaire, resulting in seventy complete scoring records. The statistical analysis was designed to examine score distribution, internal consistency, inter-item associations, inter-rater reliability, report-level stability, and the influence of clinical experience. The distribution of mean scores across the sixteen evaluation dimensions demonstrates generally favorable clinician ratings. As illustrated in Fig. 5A, average scores ranged from 3.97 to 4.49 on a five-point scale. The highest rating was observed for Q6, whereas Q9 and Q16 exhibited the lowest mean values, though still near the threshold of clinical acceptability. The ninety-five percent confidence intervals were narrow across all dimensions, indicating stable estimation and limited dispersion among evaluators. This pattern suggests that the AI-generated reports were perceived as structurally sound, terminologically standardized, and diagnostically coherent across domains.Fig. 5Comprehensive Statistical Validation of Clinician Evaluation of AI-Generated Super-Resolution Ultrasound Reports.A Overall mean scores for all sixteen evaluation dimensions across seventy rating records, displayed with ninety-five percent confidence intervals to illustrate central tendency and estimation precision. B Internal consistency analysis of the sixteen-item questionnaire using Cronbach’s alpha, including item-deleted alpha values to assess the contribution of each dimension to overall reliability. C Dimension-specific evaluation profiles with mean scores and corresponding confidence intervals, enabling direct comparison of relative strengths and variability across structural, terminological, and diagnostic domains. D Pearson correlation matrix among all questionnaire items, demonstrating inter-item associations and overall construct coherence of the evaluation framework. E Multivariate regression analysis illustrating the relative contribution of Q1 through Q15 to overall satisfaction measured by Q16, reflecting the influence of structural and content-related factors on global acceptance. F Inter-rater reliability stratified by clinical experience level, presented as intraclass correlation coefficients to quantify agreement within junior, intermediate, and senior clinician groups. G Fixed-effect estimates from mixed-effects modeling evaluating the impact of clinical experience level on each evaluation dimension, with statistically significant effects indicated accordingly. H Distributional comparison of scores for dimensions demonstrating significant experience-related effects, highlighting variability patterns across clinician experience levels. Q1 = Anatomical Structure Accuracy, referring to the correctness and precision of anatomical descriptions in the generated report, including spatial localization and structural delineation. Q2 = Completeness of Lesion Characterization, evaluating whether all relevant pathological features, boundaries, and internal characteristics are adequately described. Q3 = Reliability of Quantitative Measurements, assessing the correctness, consistency, and clinical plausibility of numerical measurements reported in the document. Q4 = Rationality of Diagnostic Interpretation, reflecting whether the diagnostic reasoning logically follows from the imaging findings and remains clinically sound. Q5 = Assessment of Omitted Information, determining whether any clinically relevant findings, anatomical structures, or contextual details are missing. Q6 = Structural Integrity of the Report, evaluating whether the report follows a coherent structure, including appropriate section organization and logical sequencing. Q7 = Standardization of Terminology, assessing the use of consistent, guideline-conformant medical terminology and avoidance of ambiguous expressions. Q8 = Formatting Consistency, reflecting adherence to standardized reporting format, including headings, units, and layout conventions. Q9 = Clarity of Expression, evaluating readability, linguistic fluency, and absence of redundant or confusing statements. Q10 = Clinical Relevance of Content, assessing whether the reported findings are appropriately focused on clinically meaningful information. Q11 = Logical Coherence Between Sections, determining whether findings, impressions, and recommendations are internally consistent. Q12 = Adequacy of Diagnostic Conclusion, evaluating whether the final impression sufficiently summarizes the imaging findings and supports clinical decision-making. Q13 = Appropriateness of Recommendations, assessing whether follow-up suggestions or clinical advice are reasonable and aligned with imaging results. Q14 = Overall Professionalism of the Report, reflecting whether the document meets standards expected in routine clinical practice. Q15 = Completeness of Report Content, evaluating whether all essential components of a standardized ultrasound report are present. Q16 = Overall Satisfaction, representing the global subjective evaluation of report quality and overall clinical acceptability. Comprehensive Statistical Validation of Clinician Evaluation of AI-Generated Super-Resolution Ultrasound Reports. A Overall mean scores for all sixteen evaluation dimensions across seventy rating records, displayed with ninety-five percent confidence intervals to illustrate central tendency and estimation precision. B Internal consistency analysis of the sixteen-item questionnaire using Cronbach’s alpha, including item-deleted alpha values to assess the contribution of each dimension to overall reliability. C Dimension-specific evaluation profiles with mean scores and corresponding confidence intervals, enabling direct comparison of relative strengths and variability across structural, terminological, and diagnostic domains. D Pearson correlation matrix among all questionnaire items, demonstrating inter-item associations and overall construct coherence of the evaluation framework. E Multivariate regression analysis illustrating the relative contribution of Q1 through Q15 to overall satisfaction measured by Q16, reflecting the influence of structural and content-related factors on global acceptance. F Inter-rater reliability stratified by clinical experience level, presented as intraclass correlation coefficients to quantify agreement within junior, intermediate, and senior clinician groups. G Fixed-effect estimates from mixed-effects modeling evaluating the impact of clinical experience level on each evaluation dimension, with statistically significant effects indicated accordingly. H Distributional comparison of scores for dimensions demonstrating significant experience-related effects, highlighting variability patterns across clinician experience levels. Q1 = Anatomical Structure Accuracy, referring to the correctness and precision of anatomical descriptions in the generated report, including spatial localization and structural delineation. Q2 = Completeness of Lesion Characterization, evaluating whether all relevant pathological features, boundaries, and internal characteristics are adequately described. Q3 = Reliability of Quantitative Measurements, assessing the correctness, consistency, and clinical plausibility of numerical measurements reported in the document. Q4 = Rationality of Diagnostic Interpretation, reflecting whether the diagnostic reasoning logically follows from the imaging findings and remains clinically sound. Q5 = Assessment of Omitted Information, determining whether any clinically relevant findings, anatomical structures, or contextual details are missing. Q6 = Structural Integrity of the Report, evaluating whether the report follows a coherent structure, including appropriate section organization and logical sequencing. Q7 = Standardization of Terminology, assessing the use of consistent, guideline-conformant medical terminology and avoidance of ambiguous expressions. Q8 = Formatting Consistency, reflecting adherence to standardized reporting format, including headings, units, and layout conventions. Q9 = Clarity of Expression, evaluating readability, linguistic fluency, and absence of redundant or confusing statements. Q10 = Clinical Relevance of Content, assessing whether the reported findings are appropriately focused on clinically meaningful information. Q11 = Logical Coherence Between Sections, determining whether findings, impressions, and recommendations are internally consistent. Q12 = Adequacy of Diagnostic Conclusion, evaluating whether the final impression sufficiently summarizes the imaging findings and supports clinical decision-making. Q13 = Appropriateness of Recommendations, assessing whether follow-up suggestions or clinical advice are reasonable and aligned with imaging results. Q14 = Overall Professionalism of the Report, reflecting whether the document meets standards expected in routine clinical practice. Q15 = Completeness of Report Content, evaluating whether all essential components of a standardized ultrasound report are present. Q16 = Overall Satisfaction, representing the global subjective evaluation of report quality and overall clinical acceptability. Internal consistency analysis further supports the robustness of the evaluation framework. The overall Cronbach’s alpha reached 0.9730, reflecting high internal consistency. As shown in Fig. 5B, item-deleted alpha coefficients remained tightly clustered between 0.9702 and 0.9723. The minimal fluctuation after removal of any single dimension indicates that no individual item disproportionately influences overall reliability and that the questionnaire measures a unified construct of report quality. A complementary visualization of the evaluation profile is presented in Fig. 5C, where mean scores and corresponding confidence intervals for each dimension are displayed simultaneously. Structural completeness, terminology standardization, and report formatting consistently achieved higher ratings, whereas dimensions related to detailed interpretative nuance demonstrated slightly greater variability. Nevertheless, all domains remained within a narrow scoring range, reinforcing the stability of clinician perception. The correlation structure among questionnaire items provides additional insight into the coherence of the instrument. The Pearson correlation matrix shown in Fig. 5D reveals positive and predominantly strong inter-item associations, with coefficients ranging from 0.54 to 0.89. Particularly strong relationships were observed among structural integrity, clarity, and diagnostic reasoning dimensions. This high degree of shared variance aligns with the elevated Cronbach’s alpha. It also supports the interpretation that the evaluation items collectively reflect an integrated perception of overall report quality rather than independent, isolated attributes. To determine whether variability originated from report content or evaluator subjectivity, report-level stability was assessed using one-way analysis of variance across the five independently generated reports. No statistically significant differences were identified for any of the sixteen dimensions, with all p-values exceeding 0.05. This finding suggests limited variability across generated reports. The observed variation is primarily attributable to differences among raters rather than fluctuations in algorithmic performance. The relative contribution of individual dimensions to overall satisfaction was explored through multivariate regression analysis, with Q16 serving as the dependent variable. As depicted in Fig. 5E, structural completeness demonstrated the strongest positive association with overall satisfaction, followed by several clarity- and content-related domains. Given the substantial inter-item correlations demonstrated earlier, multicollinearity likely influenced coefficient magnitude and direction. Therefore, these regression results could be interpreted as descriptive indicators of relative influence rather than causal determinants. Nevertheless, the analysis highlights that structural and clarity-related dimensions contribute to global clinician satisfaction. Inter-rater reliability was evaluated using intraclass correlation coefficients and further stratified by clinical experience level. The overall mean ICC was 0.3729, indicating moderate agreement across all evaluators. A clear experience-dependent pattern emerged, as visualized in Fig. 5F. Junior clinicians demonstrated minimal agreement, whereas intermediate and senior clinicians exhibited higher concordance, with ICC values exceeding 0.69. This stratification suggests that interpretive consistency increases with clinical experience and that variability among less experienced raters contributes to the moderate overall ICC. To further investigate the influence of experience on specific evaluation domains, a mixed-effects model was employed. Significant fixed effects of experience level were identified for Q2, Q5, Q6, and Q13, as summarized in Fig. 5G. Distributional analyses of these four dimensions, shown in Fig. 5H, reveal that intermediate and senior clinicians tended to provide more consistent and slightly higher evaluations, whereas junior clinicians displayed greater dispersion. These findings indicate that experience shapes the assessment of certain structural and content-related attributes while leaving overall acceptance patterns largely consistent. Taken together, the statistical analyses demonstrate excellent internal consistency of the evaluation instrument, strong coherence among questionnaire items, stable performance across independently generated reports, and experience-dependent inter-rater agreement patterns. By incorporating confidence intervals, intraclass correlation coefficients, report-level variance testing, multivariate regression, and mixed-effects modeling, the present validation addresses concerns regarding the absence of objective reliability metrics. Although the clinician sample remains limited in size, the convergence of multiple statistical indicators provides preliminary evidence supporting reliability and clinician-rated acceptability. Discussion This section is organized to clarify the logical progression of the study. We first summarize the principal findings and then confirm the technical reliability of the underlying super-resolution reconstruction framework. Subsequently, we describe the architecture of the LLM-assisted system and its hybrid optimization strategy, followed by the adaptive microbubble filtration mechanism based on the Microbubble Similarity Score (MSS). We then present comparative results against expert-driven parameter tuning, discuss clinical workflow integration and validation outcomes, and extend the application to additional vascular scenarios. Associations between super-resolution metrics and metabolic biomarkers are examined to provide a pathophysiological context. Finally, we address acoustic safety considerations and outline the principal limitations and directions for future investigation. Our principal findings are that we developed and validated an integrated large language model (LLM)–driven workflow for automated super-resolution ultrasound imaging and structured clinical report generation. The proposed system enables clinician voice commands to be translated into executable imaging parameters, dynamically optimizes microbubble filtration through the Microbubble Similarity Score (MSS), performs super-resolution reconstruction, and generates standardized diagnostic reports within approximately 4 min. Our results demonstrate that this end-to-end framework reduces manual parameter tuning, maintains high image quality, and preserves diagnostic interpretability, thereby addressing key barriers to clinical translation of super-resolution ultrasound imaging or SRUI1–4,17. The central contribution of the study is the development of a multimodal artificial intelligence framework18–21 that integrates super-resolution ultrasound imaging with large language models in order to address three key translational barriers: complex parameter optimization during acquisition, labor-intensive post-processing, and the lack of standardized and clinically contextualized interpretation and reporting22–26. Next, we performed validation of our super-resolution Reconstruction framework. First of all, it is important to confirm that the underlying localization algorithm maintains established sub-pixel performance. First, regarding microbubble localization precision, we implemented a dedicated MATLAB R2024b simulation framework to quantitatively assess sub-pixel accuracy. In this validation design, 100 synthetic microbubbles were generated within a 200 × 200 pixel field. Each simulated microbubble incorporated realistic point spread effects, and clustered spatial configurations were intentionally introduced to challenge localization robustness under overlapping conditions. Ground truth coordinates were explicitly recorded in a reference matrix prior to reconstruction. Sub-pixel localization accuracy was then quantified by computing the Euclidean pixel-wise error between detected coordinates and the corresponding ground truth positions. Visual error analysis demonstrated that the deconvolution-based localization approach improved separation of clustered microbubbles and reduced spatial overlap artifacts compared with conventional processing. Quantitative results are summarized in the Table S1. These results confirm that the proposed localization method achieves sub-pixel precision. The mean error below one pixel and the reduced median error distribution demonstrate consistent spatial refinement beyond the diffraction-limited resolution of the imaging system. Previously published in vivo validations from our study demonstrated spatial resolutions ranging from approximately 20 to 80 micrometers across multiple organs27–32, consistent with other SRUI benchmarks4,33–35. These findings confirm the stability of the underlying reconstruction pipeline upon which the present LLM-assisted framework is built. While the present study emphasizes the role of large language models in assisting clinicians with parameter configuration rather than introducing a novel localization algorithm per se, we are currently constructing gold-standard microfluidic phantoms with complex vascular geometries to further quantify imaging performance improvements in a controlled experimental environment. Then, we will discuss our LLM-assisted parameter translation and hybrid optimization architecture. Parallel to these developments in imaging, large language models (LLMs) have emerged as powerful tools for clinical decision support21,22,36–40, demonstrating capabilities in natural language understanding, knowledge synthesis, and complex reasoning. However, their application in medical imaging has been mainly restricted to post hoc report generation or static image interpretation15,20,41–43. The integration of LLMs with dynamic, real-time imaging modalities—especially those requiring adaptive parameter optimization like SRUI—remains an underexplored frontier. Comprehensive diagnostic interpretation necessitates synthesizing findings across multiple imaging modalities, correlating observations with the broader clinical context, including patient history and comorbidities, and communicating integrated insights through structured reports44–46. Addressing these interconnected challenges, this study establishes an intelligent symbiosis between our custom SRUI platform and multimodal LLMs (DeepSeek-R1 and MiniCPM-V), establishing a voice-commanded, end-to-end framework for SRUI acquisition, contextual analysis, and reporting. Now we will move to the semantic-level parameter translation. LLMs enable clinically meaningful parameter translation from complex natural language instructions into technically precise super-resolution imaging configurations. Super-resolution ultrasound reconstruction requires the selection of multiple interdependent parameters, including deconvolution kernel size, band-pass filtering coefficients, motion correction thresholds, temporal windows of interest, and localization settings. These parameters are not merely syntactic tokens but reflect domain-specific imaging knowledge and contextual clinical intent. A conventional rule-based system or lightweight natural language model typically relies on predefined keyword-to-parameter mappings. Such approaches can parse explicit time intervals or anatomical labels but cannot robustly infer implicit clinical intent, adapt parameter ranges according to image content, or resolve ambiguous instructions. In contrast, the LLM in our framework performs semantic-level reasoning. It interprets clinician instructions in context, estimates imaging-relevant quantities such as approximate microbubble size and motion severity, and configures reconstruction parameters accordingly. This capability substantially reduces the cognitive and technical burden on clinicians, particularly those without extensive experience in super-resolution imaging. In this sense, the LLM acts not merely as a parser but as an intelligent intermediary that bridges clinical language and advanced image reconstruction physics. This level of adaptive parameter generation cannot be reliably achieved using static rule-based systems. We performed dual-path optimization and cross-validation. Even for parameters that could theoretically be optimized using conventional numerical methods based on the Microbubble Similarity Score or other objective functions, the LLM provides independent estimation that enhances robustness through cross-validation. For example, adaptive microbubble filtering coefficients, microbubble concentration estimation, and MSS-related thresholding can be derived algorithmically from image statistics. However, in our framework, the LLM independently estimates microbubble concentration and image characteristics through semantic and visual reasoning. These LLM-derived estimates are then combined with conventional algorithmic outputs using weighted averaging. This dual-path estimation strategy reduces susceptibility to bias or instability arising from a single optimization pathway and improves the reliability of automatic parameter configuration. Therefore, the contribution of the LLM is not limited to replacing traditional optimization but rather augmenting it. The LLM provides contextual interpretation, cross-modal reasoning, and independent parameter estimation, which are subsequently integrated with deterministic image-processing modules. This hybrid design strengthens both automation and stability. Our integrated system overcomes fundamental limitations of conventional SRUI through a multi-layered AI architecture that redefines the clinician-machine interface. At its core, the framework leverages LLMs as “intelligent operators” capable of translating clinical intent into precise technical execution. This capability yields three key advances. The most formidable barrier—expert-dependent parameter optimization—is resolved by DeepSeek-R1’s ability to dynamically interpret natural language commands into executable protocols. When clinicians specify temporal windows (e.g., “Analyze the late washout phase from 3 to 5 min post-injection”), the LLM performs sophisticated temporal reasoning: converting relative phases into absolute timeframes using contrast injection timestamps, segmenting discontinuous intervals, and validating feasibility against acquisition duration. It automates microbubble filtration—the most technically nuanced step in SRUI reconstruction. This initiates a self-optimizing pipeline where the LLM quantifies microbubble density per frame, constructs spatial distribution maps, and calculates temporal fluctuation profiles. Leveraging these analytics, the system dynamically adjusts filtration thresholds (e.g., stricter thresholds [0.28–0.32] during high-concentration phases to suppress blooming artifacts; relaxed thresholds [0.35–0.40] during sparse phases to retain signal integrity). Closed-loop validation via the Microbubble Similarity Score (MSS > 750/1000), which serves as a standardized quality metric, ensures artifact suppression while preserving true signatures, achieving performance parity with expert manual tuning. The second innovation for clinical translation is the framework’s ability to synthesize multimodal data within the patient’s unique clinical narrative. The LLM contextualizes microvascular architecture by correlating quantitative features with relevant patient history. It recognizes elevated vascular tortuosity (13.33 ± 2.15) not merely as a statistical outlier but as a biomarker of microvascular remodeling; interprets heterogeneous perfusion via velocity vector analysis; and synthesizes these phenotypes with patient-specific factors (e.g., HbA1c 8.5%, neurocognitive symptoms). This process mirrors the clinician’s cognitive integration of imaging findings with the patient’s broader clinical environment—including prior investigations and comorbidities—to derive diagnostic meaning. Consequently, the system generates nuanced interpretations previously achievable only by subspecialists, such as: “Moderately elevated vascular density (0.7257) and tortuosity (13.33) suggest compensatory angiogenesis in the context of chronic hyperglycemia, consistent with early diabetic microangiopathy.” This effectively transforms raw data into clinically actionable intelligence grounded in holistic patient assessment. Thirdly, the system eliminates workflow fragmentation by integrating acquisition, quantification, and reporting into a seamless 4-minute process. A MATLAB-based templating system auto-populates patient demographics, scanner parameters, annotated SRUI images, quantitative metrics, and diagnostic impressions into structured reports. Significantly, these reports systematically incorporate key elements of diagnostic synthesis—relevant clinical history, correlation with prior/existing imaging findings, and contextualized interpretations—to communicate integrated insights effectively. This achieves rigorous standardization while allowing natural-language customization. Next, we will turn insights on comparative evaluation against expert-driven parameter tuning. We directly compared super-resolution reconstructions generated using parameters manually selected by an experienced clinician with those generated using the LLM-driven parameter adaptation strategy. A senior ultrasound physician with five years of experience in super-resolution ultrasound imaging first determined the core reconstruction parameters according to standard practice. The fixed SRUI reconstruction program then performed super-resolution processing based on these manually specified parameters. For the same raw contrast-enhanced ultrasound datasets, we subsequently applied the LLM-based method described in this study to automatically select the required parameters, and reconstruction was performed under identical computational conditions. Visual comparison indicates that the LLM-adapted reconstructions exhibit reduced background noise, improved delineation of microvascular structures, and improved vascular continuity (see Fig. S1). Fine vascular branches appear more clearly defined, and fragmentation artifacts commonly associated with suboptimal filtering or motion correction are attenuated. In regions with dense microbubble activity, the adaptive parameter selection mechanism appears to better balance cluster suppression and structural preservation, resulting in more coherent vascular morphology. These findings suggest that the contribution of the proposed framework extends beyond workflow automation. By translating natural language instructions and image-derived features into structured reconstruction parameters, the LLM effectively replaces manual parameter tuning without degrading imaging quality. On the contrary, in the evaluated cases, automated parameter adaptation achieved image characteristics comparable to or visually improved over experienced manual adjustment, while eliminating operator-dependent variability. This supports the role of the LLM as an intelligent parameter selection module embedded within a fixed and validated super-resolution reconstruction pipeline. Clinical validation across 14 specialists revealed profound efficiency gains and diagnostic utility. The findings of ICC (Fig. 5C) indicate that agreement increases with clinical experience. Intermediate and senior physicians achieved moderate to good reliability, whereas junior physicians demonstrated minimal agreement, suggesting that variability is largely attributable to evaluator experience rather than report inconsistency. In brief, our study included quantitative sub-pixel localization validation with ground truth simulation, contextual resolution benchmarking, expanded clinician sample size, structured multidimensional evaluation across 16 metrics, excellent internal consistency with α = 0.973, stratified inter-rater reliability analysis using intraclass correlation coefficients, and one-way analysis of variance confirming report stability. In addition, an extra analysis correlating super-resolution–derived hippocampal microvascular metrics from 32 human subjects with established clinical and metabolic parameters was performed for harder clinical anchors. Correlation analysis was conducted using Spearman correlation coefficients, and principal component analysis was applied to explore multivariate structure, as shown in the Fig. S2. Mean hippocampal blood flow velocity was positively correlated with 120-minute postprandial glucose levels, as seen in the Fig. S3). This relationship suggests compensatory hyperperfusion within hippocampal microvessels following glucose loading, consistent with cerebrovascular autoregulatory mechanisms under hyperglycemic stress47. In parallel analyses, elevated glucose levels were associated with reduced small-vessel diameter, and increased flow velocity likely represents a compensatory response to preserve cerebral perfusion47–51. Maximum blood flow velocity demonstrated significant positive associations with low-density lipoprotein cholesterol, total cholesterol, and 120-minute postprandial glucose (Fig. S4). These findings indicate potential interactions between lipid metabolism and glucose dysregulation in modulating peak hippocampal hemodynamic responses, possibly mediated by endothelial dysfunction pathways52–54. Mean vessel diameter was positively correlated with fasting glucose (Fig. S5). This observation may reflect early adaptive vascular remodeling in response to chronic hyperglycemia-induced hemodynamic stress or altered vasoactive signaling55,56. Potential mechanisms include changes in vessel wall compliance related to advanced glycation processes or hyperglycemia-associated vasodilatory effects. Nevertheless, prolonged hyperglycemia is well recognized to result in microvascular rarefaction and dysfunction. Although not statistically significant in the present cohort, small hippocampal microvessels with diameter less than 2 mm showed consistent negative correlations with HbA1c, glycated albumin, and 120-minute postprandial glucose, with correlation coefficients of −0.283, −0.263, and −0.128, respectively. The consistent directionality suggests that smaller vessels may be particularly vulnerable to chronic glycemic toxicity57–61. This hypothesis warrants validation in larger multicenter cohorts. Taken together, these objective biomarker correlations demonstrate that super-resolution imaging outputs are not limited to subjective interpretability but are quantitatively linked to established metabolic indicators. This provides preliminary evidence of pathophysiological relevance and potential translational value, including implications for individualized glycemic management and vascular risk stratification62–65. Moreover, the extension of neovascularization (Fig. 4) within carotid plaques expands the application domain from intracranial microvasculature to extracranial vascular pathology and demonstrates the adaptability of the system to atherosclerotic disease imaging. The added use case is clinically significant, as intraplaque neovascularization is closely associated with plaque vulnerability and stroke risk stratification66–70. Moreover, carotid plaque imaging introduces physiological challenges distinct from hippocampal imaging, including pronounced arterial pulsation and vessel wall motion, thereby further evaluating the robustness and generalizability of the voice-controlled super-resolution imaging and reporting framework under dynamic vascular conditions. Ensuring acoustic safety and regulatory compliance is fundamental to the clinical translation of any AI-assisted ultrasound system. In the present study, all imaging procedures were conducted under low mechanical index conditions in accordance with established clinical guidelines for contrast-enhanced ultrasound. Current Chinese clinical guidelines recommend that contrast examinations be performed at a mechanical index less than or equal to 0.3 to minimize microbubble destruction and optimize patient safety. International regulatory frameworks, including diagnostic ultrasound output standards issued by the United States Food and Drug Administration and the International Transcranial Ultrasonic Stimulation Safety Standards consensus, define the upper safety limit for non-ophthalmic diagnostic ultrasound as a mechanical index below 1.971,72. All experiments in this study were strictly performed with a mechanical index below 0.3, consistent with routine neurovascular contrast-enhanced imaging practice and substantially lower than both domestic and international regulatory upper thresholds73. Beyond adherence to recommended operating ranges, the system incorporates a hierarchical parameter governance architecture that prevents unsafe acoustic output, even in the event of erroneous or ambiguous AI interpretation. The large language model does not directly control acoustic emission. Rather, it generates parameter suggestions that are subsequently evaluated and constrained by multiple hardware and software control layers. At the device level, an independent hardware safety lock enforces a preset upper boundary on acoustic output. Within the current platform, the scanning control software imposes a maximum mechanical index of 0.4 at the device control layer. This value is embedded in the system configuration and cannot be overridden programmatically. Consequently, even if an AI-generated recommendation were to exceed safe diagnostic limits, the hardware control layer would automatically cap the executed value within the predefined boundary. Despite the encouraging results, several limitations merit consideration. First, transcranial acoustic window quality remains a fundamental constraint. In patients with excessive skull thickness or inadequate transtemporal windows, sufficient B-mode visualization of midbrain structures cannot be achieved, thereby precluding reliable contrast-enhanced acquisition and subsequent super-resolution reconstruction. In such cases, examinations were conservatively excluded rather than processed under suboptimal signal conditions. Although emerging low-frequency transducers may improve skull penetration, current performance remains partially hardware-dependent and therefore not universally generalizable. Second, patient motion introduces additional variability. While motion correction was implemented using B-mode–derived displacement matrices, excessive head movement during contrast-enhanced acquisition can exceed the compensatory limits of post-processing. Sequences with substantial motion artifacts were excluded to preserve reconstruction fidelity, reflecting a practical boundary in real-world clinical deployment. Third, super-resolution reconstruction relies on stable and sufficiently sparse microbubble distributions. To minimize clustering and signal saturation, contrast agents were administered via infusion pump and diluted to one-fifth of the original concentration. Nevertheless, variability in injection dynamics or patient-specific hemodynamics may influence microbubble density and tracking stability. Segments with excessive microbubble overlap were excluded to maintain adherence to the sparsity assumption underlying localization-based super-resolution imaging. Additional constraints include potential reductions in diagnostic accuracy in pathological states that alter vascular permeability or microbubble kinetics, which may affect signal behavior and parameter interpretation. A further limitation relates to the temporal resolution of the contrast-enhanced ultrasound acquisition. In the present study, most datasets were acquired at conventional clinical frame rates of 24–27 frames per second. We acknowledge that higher frame rates may yield improved super-resolution reconstruction quality (Fig. S7). Although this range reflects routine real-world practice and enhances clinical feasibility, it is suboptimal for localization-based super-resolution imaging. Lower temporal sampling increases the likelihood of microbubble overlap between frames, reduces tracking continuity, and may degrade reconstruction fidelity, particularly in regions with relatively fast flow. As a result, image quality under these conditions may not fully reflect the theoretical performance ceiling of the proposed framework. Future work will focus on improving robustness through hardware optimization, adaptive motion compensation strategies, and kinetic modeling tailored to altered vascular permeability states. We also plan to integrate cross-modal imaging data, including MRI and PET, and incorporate longitudinal follow-up information to enhance clinical contextualization and emulate the comprehensive synthesis performed by experienced clinicians. Under the current sample size and number of reports, senior physicians showed slightly higher mean scores than junior and intermediate physicians. Several subitems demonstrated trend-level differences, suggesting that experience may influence certain qualitative aspects of evaluation, but the statistical evidence remains limited and should be interpreted with caution. Importantly, prospective multicenter studies across different scanners, vendors, and operator experience levels will be conducted to rigorously evaluate reproducibility, external validity, and real-world generalizability. To sum up, this study presents a novel, integrated framework designed to streamline voice-commanded super-resolution ultrasound imaging acquisition, analysis, and diagnostic reporting. This integrated approach directly addresses fundamental barriers hindering broader SRUI adoption—namely operator dependency, interpretive subjectivity, and workflow fragmentation—by utilizing multimodal large language models (LLMs) to translate clinical intent into precise technical execution. Validation of the system confirms its capacity to achieve high-level parameter optimization and generate detailed diagnostic reports with efficiency gains. Critically, these reports synthesize findings from multimodal imaging, correlate observations with the patient’s clinical history, and communicate integrated insights, thereby reflecting the holistic interpretation essential for accurate diagnosis. Demonstrating its clinical readiness, the system received high utility ratings from clinicians in radiology, endocrinology, and neurosurgery departments. While challenges related to anatomical generalization and computational optimization persist, this research nonetheless establishes a new framework for integrating advanced supre-resolution imaging with rich clinical context. As conversational AI emerges as the primary interface for complex medical technology, the work described here provides a clinically informed framework for accessible point-of-care super-resolution diagnostics. Future efforts will focus on expanding this framework to multiple organs, conducting multicenter diagnostic validation, enhancing multimodal contextualization capabilities, and evolving autonomous systems capable of iterative learning from clinical interactions. Methods Study population, ethics, and safety architecture The study cohort consisted of 32 male inpatients diagnosed with type 2 diabetes according to the WHO 1999 criteria, recruited from the Endocrine and Metabolism Department of Shanghai Sixth People’s Hospital. Participants had a mean age of 57.6 ± 8.4 years. Individuals were excluded if they had experienced acute cardiovascular events, acute heart failure, NYHA class III/IV heart failure, or severe arrhythmias within the preceding six months; presented with severe hepatic or renal impairment; had documented histories of psychiatric disorders, malignancy, or substance abuse. This study was conducted in accordance with the Declaration of Helsinki. The study protocol was approved by the Ethics Committee of Shanghai Sixth People’s Hospital Affiliated to Shanghai Jiao Tong University School of Medicine under approval number 2020-236. Written informed consent was obtained from all individual participants included in the study prior to enrolment. The trial was registered with the Chinese Clinical Trial Registry (ChiCTR), a WHO Primary Registry, under the registration number ChiCTR2100048361 with registration date of July 6, 2021. The full protocol and registration details are publicly accessible at https://www.chictr.org.cn/showproj.aspx?proj=129137. First, access to the super-resolution imaging control platform is strictly restricted through an authenticated login system. Each clinical user is assigned a unique account and password, and only authorized personnel are permitted to access the system. Role-based access control is implemented to ensure that system operation, data processing, and report generation are limited to credentialed users. This mechanism prevents unauthorized access to imaging data and associated clinical information. Second, all large models used in our framework, including DeepSeek-R1 and MiniCPM, are deployed locally on dedicated institutional workstations. No patient data is transmitted to external servers or cloud-based platforms during inference. These workstations are configured in accordance with the institutional information security policies and are equipped with encryption software provided by the hospital information center. The devices operate exclusively within the institution’s internal network environment and are physically isolated from external internet access. This closed-network architecture minimizes the risk of data leakage through online transmission. In addition, external data transfer is tightly controlled. Only institutionally authorized removable storage devices, such as encrypted USB drives or external hard disks registered with the hospital information center, can be recognized by the workstation. Unauthorized storage devices are automatically blocked by the security system. This hardware-level restriction further minimizes the possibility of unauthorized data extraction. Through multi-layered protection measures, including authenticated access control, fully local model deployment, encrypted internal network operation, and restricted peripheral device access, the proposed system is designed to ensure that patient imaging data and sensitive clinical information remain securely confined within the institutional environment during the entire inference and reconstruction process. These safeguards are consistent with institutional data protection regulations and standard clinical information security practices. Speech acquisition and recognition module To support natural language interaction, we implemented a speech acquisition and recognition module that enables clinicians to issue voice commands for super-resolution ultrasound analysis. In the current study, participating clinicians were instructed to use standard Mandarin when providing spoken instructions. Because the clinicians were recruited from different geographic regions of China, natural regional accent variations were inherently present. This design allowed us to partially evaluate robustness to accent diversity under realistic clinical conditions. All audio recordings were transcribed using the FunASR toolkit (https://github.com/modelscope/FunASR). Super-resolution ultrasound imaging platform Contrast-enhanced ultrasound (CEUS) data acquisition was performed using a clinical ultrasonography system (Mindray Resona R9, Shenzhen Mindray Bio-Medical Electronics Co., Ltd., PRC) equipped with an SP5-1U phased-array transducer operating at a central frequency of 2.5 MHz. Microbubble contrast agent (SonoVue®, Bracco Imaging B.V., Switzerland) was administered via intravenous bolus injection into the cubital vein; each dose consisted of 1.6 mL SonoVue diluted to a total volume of 5 mL with saline. Transcranial imaging was conducted through the transtemporal acoustic window. Raw data were continuously acquired at a frame rate of 24–27 Hz for a duration of 5–8 minutes per acquisition, with the imaging depth set to 120 mm. All acquired raw data were stored for subsequent analysis. Evaluation of Microbubble Similarity Score The Microbubble Similarity Score (MSS) is defined, in Eq. (1), as a weighted combination of two independently derived similarity measures. Specifically,1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${MSS}=0.6\times {MSS}{\rm{\_}}2D+0.4\times {MSS}{\rm{\_}}{minicpm}$$\end{document}MSS=0.6×MSS_2D+0.4×MSS_minicpm Here, MSS_2D represents a quantitative two-dimensional similarity metric computed using the MATLAB function corr2, which evaluates the normalized cross-correlation between the observed microbubble image and a standard point spread function. MSS_minicpm represents a structural quality score generated by the MiniCPM-V model based on morphological assessment. The weighting coefficients were selected to prioritize objective pixel-level similarity while preserving morphological reasoning provided by the vision-language model. The two-dimensional similarity component is defined as Eq. (2)2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${MS}{S}_{2D}={corr}2\left(A,\,B\right)\times 1000$$\end{document}MSS2D=corr2A,B×1000where A denotes the grayscale image matrix (m×n) of the observed microbubble and B (m×n) denotes a grayscale image matrix of identical dimensions corresponding to a standard point spread function. To be more specific, we have Eq. (3)3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${MSS}{\rm{\_}}2D=\frac{{\sum }_{m}{\sum }_{n}({A}_{{mn}}-\bar{A})({B}_{{mn}}-\bar{B})}{\sqrt{({{\sum }_{m}{\sum }_{n}({A}_{{mn}}-\bar{A})}^{2})(({\sum }_{m}{\sum }_{n}{({B}_{{mn}}-\bar{B})}^{2})}}\times 1000$$\end{document}MSS_2D=∑m∑n(Amn−A¯)(Bmn−B¯)(∑m∑n(Amn−A¯)2)((∑m∑n(Bmn−B¯)2)×1000 The reference point spread function is generated using the MATLAB fspecial function with the Gaussian option to simulate an ideal diffraction-limited microbubble profile. The corr2 function computes the normalized correlation between A and B, thereby quantifying the degree to which the observed microbubble conforms to the expected Gaussian spatial distribution. The second component, MSS_minicpm, is generated by the MiniCPM-V model based on three morphological factors: the ratio of the long and short axes, which reflects circularity; the directional completeness of the microbubble contour; and the spatial dispersion of the microbubble intensity distribution. The model produces a score ranging from 0 to 1000, with higher values indicating greater conformity to an ideal microbubble morphology. To validate MSS quantitatively, we conducted a correlation analysis against expert evaluations. We first generated ten simulated high-definition Airy disk images with controlled aberration severity using a MATLAB-based optical simulation script. The simulation incorporated zero-padding to enhance spatial resolution and included defocus, astigmatism, and coma components modeled through simplified Zernike polynomial terms. Aberration severity was varied linearly from minimal to severe, and corresponding visual scores were generated to simulate realistic grading conditions. These simulated microbubble-like diffraction patterns were then processed by the AI system to compute MSS values. Expert evaluation data were obtained from 44 independently collected physician reports in which each microbubble image was manually scored. For each microbubble, we computed the mean expert score across all reports, resulting in an expertMeanPerBubble vector. No missing values were present in any microbubble column, ensuring statistical robustness. We assessed the linear association between MSS and expert mean scores using the Pearson correlation coefficient (Fig. S6), as both variables are continuous and approximately linearly related. The analysis yielded a Pearson correlation coefficient of r = 0.948 with a p-value less than 10−4, indicating a strong positive correlation. Scatter plot visualization demonstrated that individual microbubble MSS values were closely aligned with the regression line derived from expert scores, supporting concordance between AI-derived similarity assessment and human evaluation. In our framework, the MSS is primarily designed to preferentially select high-quality microbubble signals to ensure structural reliability in super-resolution reconstruction. Therefore, the threshold was not determined purely by statistical optimization but by a quality-first selection principle. Specifically, an initial cutoff of MSS ≥ 750 out of 1000 was empirically defined to retain microbubbles that closely match the reference feature distribution derived from controlled phantom experiments. This value corresponds to signals exhibiting stable point spread characteristics, appropriate spatial compactness, and intensity profiles consistent with true isolated microbubbles. Because excessive stringency may lead to insufficient microbubble density for reconstruction, we implemented an adaptive relaxation mechanism. When the number of accepted microbubbles within a given acquisition segment falls below a predefined minimum density required for stable tracking and vascular continuity, the MSS threshold is gradually reduced in a controlled manner until an adequate number of candidate microbubbles is obtained. This design is intended to prioritize high-fidelity signals while maintaining reconstruction completeness when signal sparsity occurs. At the current stage, the MSS ≥ 750 threshold has been validated in hippocampal imaging and carotid plaque imaging datasets, where it produced stable reconstruction quality and robust tracking continuity. However, we acknowledge that acoustic attenuation, perfusion characteristics, and microbubble kinetics differ across organs. Integrated LLM-driven processing framework and super-resolution reconstruction pipeline An integrated large language model (LLM)-driven framework processed the acquired data. Clinicians interacted with the system using natural language voice commands, such as instructing it to generate super-resolution images of specific vasculature within a defined post-injection time window. These commands were interpreted by a multimodal LLM module (DeepSeek-R1 for text analysis and MiniCPM-V for image analysis), which executed three core functions. First, it performed temporal parameterization, converting relative time descriptors (e.g., “arterial phase”) into absolute frame indices based on the recorded contrast injection timestamp. Second, it managed adaptive microbubble filtration, dynamically adjusting signal intensity thresholds using an iteratively adjusted parameter pipeline. This pipeline quantified microbubble density per frame using spatial distribution maps, applied stricter thresholds (0.28–0.32) during high-concentration phases to suppress blooming artifacts, and relaxed thresholds (0.35–0.40) during sparse phases to preserve signal fidelity. Output quality was validated using a proprietary Microbubble Similarity Score (MSS, scale 0–1000), triggering re-analysis if the MSS fell below 750. Third, the LLM orchestrated the overall workflow, automating subsequent steps including motion correction, singular value decomposition (SVD), and region-of-interest (ROI) selection. Specifically, parameters inferred by the large language model are not directly executed. Instead, they are combined with independently computed quantitative estimates derived from the control algorithm itself. The final operational parameter is determined by a weighted averaging strategy that integrates the LLM-derived value and the programmatically computed value. This hybrid decision mechanism effectively constrains outlier outputs, mitigates the influence of occasional semantic variability, and mitigates the risk of unsafe or unreasonable parameter configurations from being applied to reconstruction. The super-resolution reconstruction pipeline commenced with motion correction, which was realized using rigid registration applied to contrast-enhanced images averaged per cardiac cycle. Microbubble localization was achieved by isolating microbubble signals through adaptive thresholding. The point spread function (PSF) was estimated within 5×5 pixel spatial regions using single-bubble templates. Sub-wavelength localization precision was attained by calculating the pixel-intensity-weighted centroid of each identified microbubble signal. Microbubble trajectories were then linked across frames using a graph-based Kalman filter algorithm, with a maximum linking distance of 6 pixels and no limit on gap frames. Finally, vascular density and velocity maps were generated by applying Gaussian filtering to the accumulated trajectories and localizations. First, regarding the natural-language understanding component, we locally deployed the DeepSeek-R1 70B model on a dedicated workstation equipped with high-performance GPUs. The model weights were obtained from the official Ollama model repository at https://ollama.com/library/deepseek-r1. Although the base DeepSeek-R1 70B model demonstrates general reasoning capability, it lacks domain-specific knowledge in medical ultrasound imaging. In preliminary testing, it failed to correctly interpret specialized terms such as Time of Interest and Jitter Score, and it frequently produced unconstrained outputs such as bounding-box formats irrelevant to the downstream super-resolution pipeline. These observations indicated that domain adaptation through fine-tuning was necessary. Our fine-tuning objective was to train a vertical-domain model capable of converting radiologists’ natural-language voice instructions into structured parameters required for super-resolution reconstruction. Specifically, the model was required to perform structured extraction of four key parameters: Time of Interest, Jitter_Score, Region of Interest, and Imaging Requirements. It also needed to normalize colloquial anatomical descriptions into standardized terminology, for example, mapping “left hippocampus” to Hippocampus, L, and mapping qualitative motion descriptions such as “completely stable” or “severely shaking” into numerical jitter scores ranging from 0 to 100. In addition, the model was required to demonstrate robustness against linguistic variability and background narrative information unrelated to imaging control. The training dataset consisted of 150 natural-language ultrasound instructions provided independently by 15 experienced ultrasound physicians. These instructions reflected realistic clinical phrasing and included examples such as requests to render microvascular flow dynamics between specific time intervals, quantify vessel diameters, or note respiratory-induced motion artifacts. For each instruction, the same physicians manually annotated the corresponding structured super-resolution parameters. For example, an instruction describing analysis of the left hippocampus between 10 and 25 s with no image shaking was annotated as Time of Interest equal to [10, 25], Jitter Score equal to 10 on a scale from 0 to 100, Region of Interest equal to Hippocampus, L, and Output Requirements equal to speed and diameter. These paired voice instructions and structured annotations formed the supervised dataset used for fine-tuning. All 150 instructions have been made publicly available to ensure reproducibility. The fine-tuning process was conducted in multiple stages. An initial attempt using 50 samples with conservative hyperparameters resulted in underfitting and hallucinated outputs. We then implemented a data engineering phase in which high-quality synthetic augmentation was generated through rule-based scripting to ensure consistent formatting of Time of Interest as a string list representation such as [10, 25], standardized region labels, and explicit jitter mapping rules including stable mapped to 0, mild shaking mapped to 30, and severe shaking mapped to 90. This resulted in a final dataset of 150 high-consistency samples. We subsequently adopted an aggressive LoRA-based parameter-efficient fine-tuning configuration to facilitate effective domain adaptation. The LoRA rank parameter was increased from 8 to 64, and the scaling factor alpha was increased from 16 to 128 to amplify the influence of domain-specific gradients. The number of training epochs was increased to 15, and the learning rate was set to 2 × 10−4 to ensure stable convergence. The LoRA adapter was applied to the attention projection layers including q_proj, k_proj, v_proj, o_proj, as well as gate_proj, up_proj, and down_proj. Mixed precision training using bfloat16 was enabled to optimize computational efficiency. The overall model architecture consisted of four layers: an input layer receiving unstructured natural language instructions, a model layer composed of DeepSeek-R1 70B integrated with the LoRA adapter, a post-processing layer implementing regular-expression extraction and strict type conversion from string to numerical or list formats, and a final output layer generating a standardized JSON object. A robust JSON parsing function was implemented to extract the first valid JSON block from model output, convert string-based time intervals into numeric lists using safe literal evaluation, and enforce integer conversion of jitter scores. Performance evaluation was conducted on 1050 independent natural-language instructions not used during training. Quantitative analysis demonstrated greater than 99% accuracy in extracting Time of Interest, Region of Interest, and Imaging Requirements. The Jitter Score demonstrated full semantic consistency across the 0 to 100 scale, accurately reflecting physician-described motion intensity. Second, the imaging system was modeled in the frequency domain using the transfer function relationship, as shown in Eq. (4):4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\rm{T}}\times {\rm{M}}={\rm{R}}$$\end{document}T×M=R In this equation, T represents the transfer function of the ultrasound imaging system, M represents the frequency-domain matrix of the microbubble object, and R represents the frequency-domain matrix of the resulting image1 Equation (1) can be rearranged as Eq. (5):5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\rm{T}}={\rm{R}}\times {{\rm{M}}}^{-1}$$\end{document}T=R×M−1 During actual acquisition, R is observable. To estimate the true frequency-domain representation of the microbubble, it is therefore necessary to determine the transfer function T of the ultrasound imaging system. If M is modeled as an impulse response function, then we have Eq. (6):6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\rm{T}}\times 1={\rm{R}}$$\end{document}T×1=R Thus, the transfer function corresponds to the system response to an impulse input. Because isolated ultrasound microbubbles strongly reflect acoustic waves and approximate impulse sources, the imaging response of a single microbubble can be used to estimate the system transfer function. To account for the spatial heterogeneity of the point spread function, the acquired image was divided into a grid of 10 regions along the depth axis and 10 regions along the lateral axis. Within each region, 20 isolated microbubble signals were averaged to estimate the local transfer function. Following deconvolution, temporal band-pass filtering was applied to suppress low-frequency tissue motion and high-frequency noise. The lower cutoff frequency was set at approximately 2 percent of the velocity of the fastest-moving speckle, and the upper cutoff was set at approximately 30 percent of that velocity. This strategy effectively isolates microbubble motion signals while minimizing background contamination. Motion correction was then performed using B-mode video sequences as structural references. Stable anatomical features were tracked frame by frame to estimate displacement fields representing respiratory motion or probe instability. The resulting motion correction matrix was applied to the corresponding contrast-enhanced sequence, thereby isolating true microbubble motion and preventing blur accumulation during super-resolution reconstruction. For microbubble localization, each candidate signal was fitted using a two-dimensional Gaussian model. The coordinates corresponding to the maximum of the fitted Gaussian surface were defined as the sub-pixel localization of the microbubble. These coordinates were accumulated across frames to reconstruct the super-resolved vascular map. To provide quantitative error analysis, we implemented a MATLAB R2024b simulation framework. In this framework, 100 synthetic microbubbles were generated within a 200 × 200 pixel field, incorporating realistic point spread characteristics and clustered configurations to challenge localization performance. Ground-truth coordinates were recorded in a matrix for direct comparison. Sub-pixel localization accuracy was computed as the Euclidean pixel-wise distance between detected coordinates and ground truth. The quantitative results are summarized in the Table S1. LLM in parameter adaptation In our system, the core super-resolution reconstruction pipeline is fixed and implemented within the SRUI control program. DeepSeek-R1 automatically generates a MATLAB subfunction (Document S3), which plays an important role in assisting the super-resolution control program by previewing raw contrast-enhanced ultrasound (CEUS) data and providing an estimate of the microbubble density. The subfunction is designed to handle several key steps, including denoising, deconvolution, bandpass filtering, and bubble detection. The reconstruction consists of four sequential modules: deconvolution, band-pass filtering, motion correction, and microbubble localization. The mathematical structure and computational procedures of these modules are predefined and remain unchanged. What the LLM contributes is not the reconstruction algorithm itself, but the automatic selection and initialization of the key parameters required by these modules. In conventional workflows, these parameters must be manually specified by the operator prior to reconstruction. However, accurate parameter tuning requires experience with super-resolution imaging and familiarity with microbubble dynamics, which many clinicians may lack due to limited exposure or time constraints. In our framework, the LLM replaces manual parameter entry by automatically determining these core parameters based on spoken clinical instructions and the raw contrast-enhanced ultrasound data. For the deconvolution stage, the MiniCPM-V model first analyzes the raw contrast-enhanced ultrasound frames and estimates the approximate diameter of individual microbubbles. Based on this estimation, the system determines the size of the two-dimensional deconvolution kernel used in the deconvolution process. Estimating this kernel size manually would require a clinician to understand the physical relationship between microbubble dimensions and the point spread function, which may be challenging for users new to super-resolution imaging. After the kernel size is determined, the super-resolution control program performs frame-by-frame deconvolution and identifies high-intensity regions corresponding to candidate microbubble locations. The program estimates the number of microbubbles in each frame. In parallel, the MiniCPM-V model independently estimates the microbubble count directly from each frame. The final estimate of the total microbubble number for a given examination is defined as the average of the counts obtained from the reconstruction control program and the MiniCPM-V model. This averaged particle count serves as a key quantitative input for subsequent parameter adaptation. It is important to note that, for objective parameters like the Time of Interest (TOI), DeepSeek-R1 consistently produced the same results across multiple runs, with no variation. For more subjective parameters, such as the assessment of image jitter, we observed approximately 10% variability, and for microbubble concentration estimation by MiniCPM, there was 15% variability. To address this, we have implemented a protective mechanism in the super-resolution control program, where the results from the LLM are averaged with the independently generated estimates, reducing potential inconsistencies.” For the band-pass filtering stage, the DeepSeek-R1 model sets a correction factor denoted as β, which is defined as Eq. (7):7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\rm{\beta }}=\frac{{\rm{Number\; of\; Particles}}}{100}$$\end{document}β=Number of Particles100 The band-pass filter is designed to remove clustered, stationary, or slowly moving tissue signals and microbubbles, as well as high-frequency flickering noise. The filter is governed by two parameters, α1 and α2. A larger α1 enhances suppression of clustered and low-velocity signals, whereas a smaller α2 enhances suppression of high-frequency flickering noise. The baseline parameters are empirically determined as Eq. (8) and Eq. (9):8\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{\rm{\alpha }}}_{1}=50\times \frac{{Frames}}{3300}$$\end{document}α1=50×Frames33009\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{\rm{\alpha }}}_{2}=1000\times \frac{{Frames}}{3300}$$\end{document}α2=1000×Frames3300where Frames represents the total number of frames in the contrast-enhanced ultrasound sequence. When β is large, indicating a high candidate microbubble count and an increased probability of clustering, the filter parameters are adjusted as Eq. (10) and Eq. (11):10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{\rm{\alpha }}}_{1}^{{\prime} }=\frac{{{\rm{\alpha }}}_{1}}{{\rm{\beta }}}$$\end{document}α1′=α1β11\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{\rm{\alpha }}}_{2}^{{\prime} }=\frac{{{\rm{\alpha }}}_{2}}{{\rm{\beta }}}$$\end{document}α2′=α2β In this way, the LLM does not alter the filtering algorithm itself, but adaptively modifies the filtering strength according to the estimated microbubble burden. For the motion correction module, DeepSeek-R1 interprets the clinician’s natural language instruction and extracts a parameter denoted as Jitter_Score_deepseek, ranging from 0 to 100, where higher values indicate more severe image motion. If Jitter_Score_deepseek equals 0, the model determines that motion correction is minimal and sets the maximum allowable shift parameter max_shift to 10. Otherwise, the maximum motion correction magnitude is defined as Eq. (12):12\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$max {{\_}}{shift}{{\_}}{deepseek}=10+90\times \frac{\,{Jitter}{{\_}}{Score}{{\_}}{deepseek}}{100}$$\end{document}max_shift_deepseek=10+90×Jitter_Score_deepseek100 Independently, the MiniCPM-V model downsamples the original sequence at 10 frames per second and computes a motion score denoted as Jitter_score_minicpm, also ranging from 0 to 100. Because motion artifacts in contrast-enhanced ultrasound are often subtle and clinicians’ verbal assessments provide valuable contextual information, we assign greater weight to the language-derived estimate. The final motion score used for reconstruction is therefore calculated, in Eq. (13), as a weighted combination:13\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${Jitter}{{\_}}{score}=max {{\_}}{shift}{{\_}}{deepseek}\times 0.8+max {{\_}}{shift}{{\_}}{minicpm}\times 0.2$$\end{document}Jitter_score=max_shift_deepseek×0.8+max_shift_minicpm×0.2 This weighted value determines the effective motion correction range applied by the fixed reconstruction module. Finally, in the microbubble localization stage, the super-resolution control program performs deterministic localization independent of the LLM. Raw microbubble intensity profiles are fitted using a two-dimensional Gaussian function, and the precise x and y coordinates corresponding to the peak of the fitted Gaussian surface are extracted. These coordinates are defined as the sub-pixel localization positions of the ultrasonic microbubbles. The fitting algorithm and localization method remain entirely fixed within the SRUI reconstruction pipeline. In our framework, microbubble velocity is computed as the first-order temporal derivative of the displacement vector. Specifically, we have Eq. (14),14\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$v=\frac{d\vec{s}}{{dt}}=\frac{\vec{P\left(t+\Delta t\right)}-\vec{P\left(t\right)}}{\Delta t}$$\end{document}v=ds⃗dt=Pt+Δt⃗−Pt⃗Δtwhere \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\vec{P\left(t\right)}$$\end{document}Pt⃗ represents the displacement vector of a microbubble at time t, and Δt denotes the frame interval. We emphasize that vector Doppler imaging was not used in this study, and this clarification has been explicitly added to the revised manuscript to avoid misinterpretation.” Automated report generation The framework automatically generated diagnostic reports. Quantitative parameters, including vascular density, tortuosity, and flow velocity, were extracted directly from the super-resolution maps. The LLM synthesized these imaging findings with the patient’s clinical history and current medical guidelines to produce structured reports. Key sections of these reports included a “Clinical Correlation” section, which interpreted findings, and “Management Recommendations,” such as advising glycemic optimization, multidisciplinary follow-up, or interval rescans. Reports were formatted into editable Microsoft Word or Excel documents using MATLAB templating functions. Validation and statistical analysis We generated five diagnostic reports and recruited a cohort of 14 independent clinician evaluators, comprising junior, intermediate, and senior physicians. Each clinician independently assessed all five reports, resulting in a total of 70 structured evaluation observations. The evaluation framework consisted of 16 structured dimensions encompassing diagnostic accuracy, completeness, clarity, standardization, and clinical usability. Assessments used Likert scales ranging from 1 to 5. Statistical analysis of the validation results, along with inter-rater variability quantification, was conducted using MATLAB’s built-in statistical functions. Computational infrastructure All data processing and analysis were executed on a high-performance workstation equipped with two AMD EPYC 7763 processors, two NVIDIA A800 graphics cards, and 128 GB of RAM. In our system, raw contrast-enhanced ultrasound data are transmitted in real time from the ultrasound platform to the workstation through a 10-gigabit network interface. However, super-resolution reconstruction is not performed simultaneously with data acquisition in a frame-by-frame streaming manner. If reconstruction were executed concurrently with acquisition, the resulting super-resolution maps may compromise reconstruction accuracy. This is because several essential steps in the pipeline, including temporal band-pass filtering and motion correction, require access to the complete time sequence rather than partial temporal segments. These operations depend on global temporal information to reliably suppress tissue motion, correct respiratory displacement, and ensure consistent microbubble localization across frames. Therefore, super-resolution reconstruction and structured report generation are initiated immediately after completion of contrast-enhanced data acquisition, rather than during acquisition itself. For a typical contrast-enhanced ultrasound sequence containing approximately 5000 frames, with each frame including roughly 30–100 microbubbles, the processing time distribution is as follows. Motion correction requires 50–90 s. Band-pass filtering requires 0.5 to 1 second. Microbubble localization requires 30–60 s. Microbubble tracking requires 60 to 70 s. Structured report generation requires 20–30 s. Based on these measurements, the total post-acquisition processing time is approximately four minutes from the completion of ultrasound imaging to the generation of the super-resolution reconstruction and diagnostic report. Reducing the acquisition frame rate proportionally reduces the total number of frames and therefore shortens reconstruction time. However, lower frame rates increase the probability of missing transient microbubble appearances and reduce trajectory continuity, which can degrade localization density and tracking robustness. Similarly, in the localization stage, omitting the deconvolution step reduces computational time but results in a measurable decline in localization precision. Supplementary information Supplementary_Materials_and_consort Supplementary_Materials_and_consort