A multimodal dataset and predictive model for the treatment of uterine fibroids with focused ultrasound ablation surgery Authors: Qianru Zeng, Ling Fan, Shuaibin Liu, Qian Wang, Yu Zhao, Min Zhou, Li Ling Published: 2026 ABSTRACT AbstractUterine fibroids are a common indication for image-guided, uterus-preserving therapies such as focused ultrasound ablation surgery (FUAS), yet substantial interpatient variability in treatment difficulty limits reproducible analysis. Here, we present a curated multimodal dataset comprising structured clinical variables, continuous multisequence magnetic resonance imaging (MRI), procedural parameters, and quantitatively defined treatment outcome indicators from patients undergoing FUAS for uterine fibroids. The dataset integrates heterogeneous data modalities within a standardized and anonymized framework to support reproducible research and downstream reuse. To facilitate technical validation, we further provide a reference multimodal prediction pipeline that integrates dynamic MRI sequences with structured clinical data to estimate treatment difficulty and derive a rule-based FUAS recommendation label. Using this baseline implementation, the dataset supports joint regression and classification tasks, achieving a mean absolute percentage error of 0.265 and an area under the receiver operating characteristic curve of 0.931 on an independent test set. This openly available dataset provides a valuable resource for benchmarking and for multimodal data fusion research related to FUAS treatment of uterine fibroids. Background & Summary Uterine fibroids are benign smooth muscle tumors of the uterus and represent the most common pelvic neoplasm among women of reproductive age, with reported prevalence estimates ranging from 20% to 80% depending on population characteristics and diagnostic criteria1,2. Symptomatic fibroids are frequently associated with heavy menstrual bleeding, pelvic pain, pressure-related symptoms involving adjacent organs, and adverse reproductive outcomes3,4, resulting in a substantial burden on quality of life and healthcare systems5,6. A broad range of therapeutic strategies are currently available for the management of uterine fibroids, including pharmacological therapies, surgical interventions, interventional radiology procedures, and image-guided physical ablation techniques7,8. Although hysterectomy provides definitive symptom resolution, it permanently eliminates reproductive potential. Myomectomy preserves the uterus but is associated with surgical morbidity and a considerable risk of recurrence9. Interventional approaches such as uterine artery embolization can effectively reduce fibroid volume; however, concerns persist regarding their potential impact on uterine perfusion and subsequent reproductive outcomes in selected patient populations10,11. Focused ultrasound ablation surgery (FUAS) is a noninvasive, image-guided thermal ablation technique that enables targeted destruction of fibroid tissue while preserving surrounding uterine structures12. Its lack of ionizing radiation, minimal invasiveness, potential for fertility preservation, and rapid postprocedural recovery have contributed to increasing clinical adoption11,13–15. FUAS has become an important option for symptomatic patients or those with fertility intentions. Nevertheless, both treatment outcomes and procedural difficulty exhibit substantial interpatient variability, even when standardized treatment protocols are applied. Reliable pretreatment assessments of treatment feasibility, procedural complexity, and expected short-term efficacy therefore remain major challenges in clinical practice. Previous studies have investigated outcome prediction for FUAS using statistical models and machine learning approaches based on limited clinical or imaging features16–18. However, the effectiveness of FUAS is influenced by a complex interplay of factors, including patient-specific characteristics, fibroid morphology, vascularity, spatial relationships to surrounding tissues, and procedural parameters16,19. Single-modality data are often insufficient to capture this complexity, and the absence of publicly available, well-annotated multimodal datasets has constrained systematic methodological development and external validation20–22. Recent advances in artificial intelligence(AI), particularly multimodal data fusion techniques, have demonstrated the potential to integrate heterogeneous data sources such as medical images and structured clinical variables23. These approaches provide a promising framework for modeling complex treatment processes and outcomes. However, progress in this area remains limited by the scarcity of standardized, high-quality datasets that combine longitudinal imaging, detailed clinical annotations, and procedural outcome measures in a reproducible and openly accessible format. Here, we present a comprehensive multimodal dataset of patients who underwent FUAS for uterine fibroids,including integrating structured clinical data, continuous pre- and posttreatment multi-sequence magnetic resonance imaging (MRI), procedural parameters, and quantitatively defined treatment outcome indicators. To facilitate downstream methodological development and benchmarking, we provide a reference multimodal prediction pipeline that demonstrates how dynamic MRI sequences and structured clinical features can jointly be utilized to estimate treatment difficulty and derive a rule-based, clinically interpretable FUAS recommendation label. This dataset is intended to be an openly accessible resource to support reproducible research, comparative evaluation of predictive methods, and future studies on data-driven analysis of FUAS treatment for uterine fibroids. Methods Ethics statement The study protocol was approved by the Research Ethics Committee of the Second Affiliated Hospital of Chongqing Medical University (Approval No.: 2025 Review (81)). Given that this work involved retrospective analysis of routinely acquired clinical data without direct patient contact, the requirement for informed consent was waived by the ethics committee. All methods were carried out in accordance with relevant institutional guidelines and regulations.All data released with this dataset were fully anonymized prior to publication to protect patient privacy, and approval for public data sharing was granted by the institutional review board. Patient cohort and data sources This dataset was retrospectively constructed using clinical and imaging records from the Focused Ultrasound Ablation Surgery (FUAS) Treatment Center for Gynecology at the Second Affiliated Hospital of Chongqing Medical University. The overall patient inclusion and exclusion workflow is illustrated in Fig. 1. Patient inclusion and exclusion workflow. This figure illustrates the screening process used to construct the dataset, including the initial patient pool, inclusion criteria, exclusion criteria, and final number of patients retained for analysis. A total of 669 consecutive patients who underwent FUAS for uterine fibroids between February 2022 and October 2025 were initially screened. Patients were included if they (i) had a confirmed diagnosis of uterine fibroids, (ii) completed pre- and posttreatment magnetic resonance imaging (MRI) according to the standardized protocol, (iii) underwent FUAS using the same therapeutic system, and (iv) had complete procedural records and posttreatment outcome assessments. Patients were excluded if key clinical variables, imaging data, or treatment outcome indicators were missing. After applying the inclusion and exclusion criteria, 308 patients with complete multimodal data were retained in the final dataset. Each patient constitutes a single dataset entry and is uniquely identified using a consecutive anonymized identifier. Data modalities and acquisition The dataset comprises three primary data modalities: structured clinical data, imaging data, and treatment-related data. All modalities are linked at the individual patient level using a unique anonymized identifier. Clinical data Structured clinical data include demographic characteristics, symptom assessments, and relevant pretreatment clinical features. Recorded variables included age (converted to 5-year age bins, e.g., “35–39”, “40–44”, for public release), body mass index (BMI), fibroid-related symptoms, and the presence or absence of abdominal wall scars within the planned treatment region. Height and weight were removed from the public dataset because they were not essential for the predictive models. Original exact ages were used only for internal validation and are not shared. Fibroid-related symptoms were quantified using standardized clinical instruments. Heavy menstrual bleeding was assessed using a five-point scale derived from the uterine fibroid symptom and quality of life (UFS-QoL) questionnaire24,25. Dysmenorrhea severity was evaluated using a 10-point visual analog scale (VAS)26. Symptoms related to compression of adjacent organs, including urinary frequency and constipation, were recorded as binary variables. The presence of anemia was determined based on routine clinical evaluation and recorded as a binary variable. The presence of abdominal wall scars intersecting the anticipated ultrasound beam path was documented as a binary feature, as this factor may influence energy ultrasound transmission during FUAS. Imaging data Imaging data include pretreatment ultrasound examinations and multi-sequence MRI acquired according to a standardized protocol. Ultrasound imaging Color Doppler ultrasound examinations were performed prior to treatment and independently reviewed by two attending sonographers with more than three years of clinical experience. The following parameters were recorded: uterine position, uterine size, fibroid type, fibroid location, fibroid dimensions, echogenicity, and vascularity grade. Magnetic resonance imaging All MRI examinations were performed using a Siemens Avanto 1.5 T scanner. The imaging protocol included T2-weighted imaging (T2WI), T1-weighted imaging (T1WI), and contrast-enhanced (CE) sequences. T2WI was acquired with a repetition time (TR) of 3770 ms and an echo time (TE) of 76 ms, with a slice thickness of 5 mm and an interslice gap of 1.25 mm. T1WI was acquired with TR = 672 ms and TE = 10 ms using the same slice parameters. Postcontrast imaging employed a TR of 4.7 ms, a TE of 1.6 ms, a slice thickness of 4 mm, and an interslice gap of 0.8 mm. Images were acquired in axial, sagittal, and coronal planes to cover the uterus and bilateral adnexal regions. The superior boundary of the scanning range included the level of the iliac vessel bifurcation. A gadolinium-based contrast agent (gadoteric acid meglumine) was administered intravenously at a dose of 0.2 ml/kg body weight, followed by dynamic contrast-enhanced axial imaging for 12 phases with a temporal resolution of 16 s. The total scan duration was approximately 20 min. MRI images were independently reviewed by two attending radiologists with more than three years of experience. Imaging-derived variables included uterine position, uterine size, fibroid type and location, fibroid volume, abdominal wall thickness, distance from the back side of uterine fibroids to the skin, distance from the dorsal side of the fibroids to the pyriformis muscle, signal intensity characteristics on T1WI and T2WI, and degree of contrast enhancement. Representative imaging features are illustrated in Fig. 2. Representative MRI-derived features included in the dataset. Representative magnetic resonance images illustrate imaging-derived measurements and categorical features recorded in the dataset. (A) (a) Abdominal wall thickness. (b) Distance from the posterior margin of uterine fibroids to the skin surface. (c) Distance from the dorsal margin of uterine fibroids to the sacrococcygeal region. (B) Uterine fibroid type. (C) Signal intensity characteristics of uterine fibroids on T2-weighted imaging. (D) Degree of contrast enhancement on contrast-enhanced MRI. All released JPEG images were resampled to a uniform spatial resolution of 224 × 224 pixels. The corresponding in-plane pixel spacing is 0.625 × 0.625 mm, which is constant across the entire dataset. Original DICOM pixel spacing was 0.625 mm; no additional scaling was applied. The publicly released dataset includes only the sagittal T2WI and CE sequences. Other acquired planes (axial, coronal) and non-contrast sequences are not part of the shared dataset to maintain a manageable data volume while preserving key diagnostic information. Treatment and procedural data All FUAS procedures were performed using a JC200 Focused Ultrasound Tumor Therapeutic System (Chongqing Haifu Technology Co., Ltd., Chongqing, China). The therapeutic ultrasound frequency was 1.0 MHz, with a focal length of 170 mm and a transducer diameter of 200 mm. Real-time lesion localization and treatment monitoring were conducted using a Mindray DC-80 ultrasound system. Procedural parameters automatically recorded by the therapeutic system included average sonication power,overall treatment time,sonication time,treatment intensity, and treatment energy. Sonication power ranged from 0 to 400 W. 400 W was used for most patients, with reductions applied intraoperatively as needed (e.g., due to treatment difficulty or patient discomfort). These parameters were exported in structured electronic format and incorporated into the dataset without post hoc modification. Treatment outcome measures and label definition Treatment efficacy and adverse events Treatment efficacy was evaluated using non-perfused volume (NPV) measurements obtained from contrast-enhanced ultrasound (using SonoVue, sulfur hexafluoride microbubbles) immediately after the procedure and from contrast-enhanced MRI performed one day post-treatment. The non-perfused volume ratio (NPVR) was calculated as follows: Higher NPVR values indicate a greater proportion of ablated fibroid tissue. Intraoperative and postoperative adverse events were recorded and graded according to the Society of Interventional Radiology (SIR) classification system27,28. Recorded events included anesthesia-related reactions, intraoperative discomfort or pain, skin injury, neurological symptoms, and postoperative complications. All adverse event data were stored as structured categorical variables. Quantification of treatment difficulty Three quantitative indicators were used to characterize treatment difficulty: sonication time, NPVR, and energy-efficiency factor (EEF). Sonication time represents the cumulative duration during which ultrasound energy was actively delivered to the fibroid. EEF reflects the energy required to ablate a unit volume of tissue and was calculated as follows: Where (η) denotes a system-specific focusing factor (set to 0.7), (P) is the sonication power (W), (t) is the sonication time (s), and (V) is the NPV (mm3). Each indicator was discretized into three ordinal levels according to predefined thresholds, resulting in scores ranging from 0 to 2 (Table 1). These scores were retained as independent labels within the dataset to support downstream modeling and comparative analysis. Table 1The discretization thresholds for the three difficulty indicators.IndicatorScore 0Score 1Score 2NPVR≥70%50–70%<50%EEF (J/mm3)≤1.61.6–5.2å 5.2Sonication time≤484.5484.5–1129.8å 1129.8Boundary conventions: “≥” includes the lower bound; “>” excludes the lower bound. These thresholds were derived from the first quartile (Q1) and third quartile (Q3) of the cohort distribution: Q1(EEF) = 1.6 J/mm3, Q3(EEF) = 5.2 J/mm3; Q1(sonication time) = 484.5 s, Q3(sonication time) = 1129.8 s. The discretization thresholds for the three difficulty indicators. Boundary conventions: “≥” includes the lower bound; “>” excludes the lower bound. These thresholds were derived from the first quartile (Q1) and third quartile (Q3) of the cohort distribution: Q1(EEF) = 1.6 J/mm3, Q3(EEF) = 5.2 J/mm3; Q1(sonication time) = 484.5 s, Q3(sonication time) = 1129.8 s. FUAS Recommendation label To generate a clinically interpretable binary label for treatment recommendation, a rule-based decision tree was constructed using three quantitative indicators: NPVR, EEF, and sonication time. In contrast to additive scoring systems, which may paradoxically classify a patient with poor ablation efficacy (NPVR < 50%) as low or high difficulty16, the decision tree applies a hierarchical, efficacy-first logic. It first screens by NPVR, then further stratifies by EEF and sonication time.The labeling rules are defined as follows: (i)First branch – efficacy criterion: If NPVR < 50%, the treatment outcome is considered poor, and the case is classified as “Not Recommended”. Cases with NPVR ≥ 50% proceed to the second branch. First branch – efficacy criterion: If NPVR < 50%, the treatment outcome is considered poor, and the case is classified as “Not Recommended”. Cases with NPVR ≥ 50% proceed to the second branch. (ii)Second branch – difficulty criterion (applies only when NPVR ≥50%): Q3 thresholds derived from the cohort are: Q3(EEF) = 5.2 J/mm3 and Q3(sonication time) = 1129.8 s. If EEF >5.2 J/mm3 and sonication time >1129.8 s, the case is classified as “Not Recommended”. Second branch – difficulty criterion (applies only when NPVR ≥50%): Q3 thresholds derived from the cohort are: Q3(EEF) = 5.2 J/mm3 and Q3(sonication time) = 1129.8 s. If EEF >5.2 J/mm3 and sonication time >1129.8 s, the case is classified as “Not Recommended”. (iii)Otherwise (i.e., either EEF ≤ 5.2 J/mm3 or sonication time ≤ 1129.8 s, or both), the case is classified as “Recommended”. Otherwise (i.e., either EEF ≤ 5.2 J/mm3 or sonication time ≤ 1129.8 s, or both), the case is classified as “Recommended”. The complete decision logic is illustrated in Fig. 3. The FUAS recommendation label is defined retrospectively based on actual post-treatment outcomes and is intended as a ground-truth label for training predictive models to estimate pre-treatment suitability for FUAS. The label, along with its constituent indicators (NPVR, EEF, sonication time), is provided as an explicit annotation field within the dataset. Rule-based decision tree for FUAS recommendation labeling. This figure illustrates the predefined decision tree used to assign the FUAS recommendation label based on three quantitative treatment difficulty indicators: NPVR, EEF, and sonication time. Each case is categorized as “Recommended” or “Not Recommended” according to the specified thresholds. Data Records The dataset described in this Data Descriptor is publicly available via the Figshare repository29 and is organized to facilitate reproducible analysis and downstream reuse. All data are fully anonymized and indexed using unique patient identifiers. The dataset integrates structured tabular data and multisequence MRI data, which are linked at the individual patient level. Dataset structure For each patient, the dataset includes (i) structured clinical and demographic variables, (ii) imaging-derived features extracted from ultrasound and MRI, (iii) procedural and treatment-related parameters, (iv) quantitative treatment outcome indicators, and (v) rule-based FUAS recommendation labels. Each patient is represented by a single anonymized identifier (ID), ranging from “001” to “308”. This identifier serves as the primary key linking all corresponding data modalities, including structured data records and imaging data (see Fig. 4). Correspondence between structured data records and MRI sequences. This schematic illustrates the one-to-one linkage between structured tabular records and corresponding MRI image sequences using unique anonymized patient identifiers. Each identifier connects clinical variables, procedural parameters, outcome indicators, and imaging data. In total, the dataset contains 19,820 MRI images consisting of continuous sagittal T2WI and CE image sequences acquired prior to treatment. Structured data files All structured data are provided in comma-separated values (CSV) format to ensure compatibility with commonly used statistical and machine learning software. Each row corresponds to a single patient, and each column corresponds to a predefined variable. A detailed data dictionary lists all variable names, descriptions, data types, allowed values, and encoding schemes. Clinical and imaging-derived variables The primary structured data file contains demographic characteristics (age group, BMI), symptom scores (heavy menstrual bleeding, dysmenorrhea, urinary frequency/constipation, anemia), ultrasound findings (uterine position, fibroid type, location, volume, vascularity), and MRI-derived measurements (signal intensity on T1WI/T2WI, degree of contrast enhancement, abdominal wall thickness, distance metrics related to fibroid location). Categorical variables are encoded using standardized numeric labels as described in the data dictionary. Treatment and procedural parameters A dedicated structured data file records FUAS procedural parameters generated automatically by the therapeutic system. Recorded variables include average sonication power, overall treatment time, sonication time, treatment intensity, and total treatment energy. All procedural parameters are stored as continuous numerical values and reflect system-generated outputs without manual postprocessing or modification. Treatment outcome indicators and labels Quantitative treatment outcome indicators are provided as explicit dataset labels: non-perfused volume ratio (NPVR) measured by posttreatment CE MRI, energy-efficiency factor (EEF), and sonication time. In addition, a binary FUAS recommendation label (“Recommended” or “Not Recommended”) is provided for each patient, derived using the rule-based decision tree described in the Methods. MRI image records Image format and organization MRI images are provided in standard JPEG format. All images have been anonymized to remove personal identifiers from both pixel data and file headers. Images are organized in a hierarchical directory structure based on patient ID and imaging sequence type. Within each subdirectory, individual image files are named according to the following convention: [PatientID]_[SequenceType]_[FrameIndex].jpeg. For example, 001_T2WI_12.jpeg denotes the 12th frame of the sagittal T2-weighted sequence for patient 001. Image quantity and correspondence Across the full dataset, a total of 19,820 MRI images are provided, consisting of continuous sagittal T2WI and CE sequences acquired prior to treatment. Each patient contributes a variable number of frames depending on the acquisition protocol. The correspondence between structured data records and MRI image sequences is maintained exclusively through the anonymized patient identifier, ensuring a one-to-one linkage between tabular data and imaging data (see Fig. 4). Data anonymization and quality control See the anonymization description above (section “Data Anonymization and Quality Control”). Quality control procedures excluded cases with missing key variables, incomplete imaging, or missing outcome records. Imaging data were reviewed to confirm correct sequence labeling. Data access and reuse considerations All data files (structured CSV files, MRI images, and code) are deposited in the Figshare repository as a single ZIP archive. The archive preserves the original folder structure and file naming convention; standard unzip software can be used to extract all contents. The dataset is intended to support methodological development, benchmarking of predictive models, and exploratory analyses related to FUAS treatment for patients with uterine fibroids. The modular organization allows users to selectively utilize specific data modalities. Users are encouraged to cite the associated Data Descriptor when the dataset is used and to comply with applicable ethical standards. Data Overview The dataset comprises 308 patients who underwent FUAS for uterine fibroids. Table 2A–C summarize the key clinical, imaging, and procedural characteristics of the cohort. Table 2(A–C) Baseline clinical, imaging, and procedural characteristics of the FUAS dataset.(A) Clinical characteristics of the FUAS cohort (n = 308)CharacteristicsDataAge, median (IQR), y40 (35–45)BMI, median (IQR), kg/m222.5 (20.9–24.8)Excessive menstruation, median (IQR)1.0 (1.0–2.0)Dysmenorrhea, median (IQR)0 (0–1.0)Frequent urination and constipation, n (%) Yes36 (11.7%) No272 (88.3%)Anemia, n (%) Yes78 (25.3%) No230 (74.7%)Abdominal surgical scar, n (%) Yes188 (61.0%) No120 (39.0%)(B) Imaging characteristics.CDFI blood flow grade of uterine fibroids, n (%) Grade 144 (14.3%) Grade 292 (29.9%) Grade 382 (26.6%) Grade 490 (29.2%)Location of the uterus, n (%) Anteverted217 (70.5%) Median15 (4.9%) Retroverted76 (24.7%)Type of uterine fibroids, n (%) Intramural193 (62.7%) Subserousal71 (23.1%) Submucosal44 (14.3%)Location of uterine fibroids, n (%) Anterior152 (49.4%) Lateral56 (18.2%) Posterior81 (26.3%) Fundus19 (6.2%)Volume of uterine fibroids, median (IQR), mm382477.3 (43898.3–141381.4)Signal intensity on T1WI, n (%) Hypointensity41 (13.3%) Isointensity245 (79.5%) Homogeneous-type hyperintense21 (6.8%) Heterogeneous-type hyperintense1 (0.3%)Signal intensity on T2WI, n (%) Hypointensity157 (51.0%) Isointensity81 (26.3%) Homogeneous-type hyperintense28 (9.1%) Heterogeneous-type hyperintense42 (13.6%)Contrast enhanced MRI, n (%) Mild53 (17.2%) Moderate150 (48.7%) Significant105 (34.1)Abdominal wall thickness, median (IQR), mm26.0 (21.0–31.75)Distance from the dorsal side of the fibroids to the pyriformis muscle, median (IQR), mm24.0 (12.0–43.0)Distance from the back side of uterine fibroids to the skin, mean ± SD, mm101.0 ± 20.6(C) Procedural and outcome characteristics.Average sonication power, median (IQR), W400.0 (400.0–400.0)*Overall treatment time, median(IQR), min70.0 (52.0–98.0)Sonication time, median (IQR), s729.5 (484.5–1129.8)Treatment intensity, mean ± SD, s/h630.3 ± 139.1Treatment energy, median (IQR), J288400.0 (188087.5–444500.0)EEF,median (IQR), J/mm32.8 (1.6–5.2)Contrast-enhanced ultrasound NPVR, mean (range), %92.3 (19.9–100.0)Post MRI NPVR, mean (range), %87.3 (19.6–100.0)This table summarizes demographic variables, symptom assessments, imaging-derived features, and treatment-related parameters for all 308 patients included in the dataset. Continuous variables are reported as means ± standard deviations or medians with interquartile ranges(IQRs), depending on their distribution. Categorical variables are presented as counts. Variable definitions and encoding schemes are provided in the accompanying data dictionary. *Actual sonication power ranged from 139 W to 400 W (mean 395 W); however, because 88.6% of patients received the maximum 400 W, the 25th, 50th, and 75th percentiles are all 400 W, giving an IQR of 400–400. (A–C) Baseline clinical, imaging, and procedural characteristics of the FUAS dataset. This table summarizes demographic variables, symptom assessments, imaging-derived features, and treatment-related parameters for all 308 patients included in the dataset. Continuous variables are reported as means ± standard deviations or medians with interquartile ranges(IQRs), depending on their distribution. Categorical variables are presented as counts. Variable definitions and encoding schemes are provided in the accompanying data dictionary. *Actual sonication power ranged from 139 W to 400 W (mean 395 W); however, because 88.6% of patients received the maximum 400 W, the 25th, 50th, and 75th percentiles are all 400 W, giving an IQR of 400–400. Based on the rule-based decision tree (see Methods), 272 patients (88.3%) were classified as “Recommended” and 36 (11.7%) as “Not Recommended”. The class imbalance should be considered when evaluating predictive models; precision-recall (PR) metrics are recommended alongside ROC-AUC. Technical Validation To assess the internal consistency, completeness, and practical usability of the dataset, a series of validation analyses were conducted. These analyses include descriptive statistical assessment of structured variables, verification of the rule-based labeling strategy, and a reference multimodal prediction experiment designed to demonstrate potential downstream applications of the dataset. Importantly, the predictive model presented here is provided solely as a baseline implementation for dataset validation and is not intended to represent an optimized or definitive clinical decision-support system. Statistical validation of structured variables Descriptive statistical analyses were performed to examine the distributional characteristics of all structured clinical, imaging-derived, and procedural variables. Normality of continuous variables was assessed using the Shapiro–Wilk test. Variables with approximately normal distributions are reported as mean ± standard deviations, whereas nonnormally distributed variables are summarized as medians with interquartile ranges. The dataset exhibits substantial heterogeneity across demographic characteristics, fibroid morphology, imaging features, and treatment parameters, supporting its suitability for developing and benchmarking predictive models. Key treatment-related indicators, including sonication time, EEF, and NPVR, span wide value ranges, indicating a broad spectrum of procedural difficulties and treatment responses (Fig. 5A–C). This variability supports the suitability of the dataset for developing and benchmarking predictive models across diverse clinical scenarios. Distribution of treatment difficulty indicators and validation of recommendation labels. (A) Box plots showing the distribution of sonication time across the full cohort. (B) Box plots showing the distribution of energy efficiency factor (EEF) across the full cohort. (C) Bar chart of non-perfused volume ratio (NPVR) categories: <50% (6, 1.9%), 50–70% (24, 7.8%), and >70% (278, 90.3%). (D) Receiver operating characteristic (ROC) curve illustrating the discriminative performance of the logistic regression model used to validate the FUAS recommendation labeling strategy. (E) Forest plot showing odds ratios (OR) and 95% confidence intervals (CI) for variables associated with the FUAS recommendation label. *p < 0.05; **p < 0.01. The x-axis includes a break to compress the scale. OR, odds ratio. Validation of rule-based FUAS recommendation labels The FUAS recommendation label was derived using a predefined, rule-based decision tree constructed from three quantitative indicators: NPVR, EEF, and sonication time. To evaluate the internal consistency of this labeling strategy, a multivariable logistic regression analysis was performed using the structured dataset. The logistic regression model demonstrated good calibration (Hosmer–Lemeshow test, P = 0.652) and satisfactory discriminative performance (accuracy 89.94%, AUC 0.902, 95% CI: 0.859–0.945, p <0.0001)(Fig. 5D). These findings indicate that the rule-based recommendation label is internally coherent with the underlying clinical and imaging-derived variables. Analysis of regression coefficients further indicated that MRI-derived features contributed meaningfully to the recommendation outcome. In particular, signal intensity characteristics on T1WI and degree of contrast enhancement were identified as statistically significant independent predictors (Fig. 5E). This observation is consistent with the clinical understanding that T1WI and contrast enhancement correlate with fibroid vascularity and ablation difficulty16,19. Multimodal consistency validation To further assess whether the released imaging and structured data modalities contain coherent and complementary information, we conducted an additional validation experiment using a reference multimodal learning framework. The purpose of this analysis was not to establish a clinically deployable prediction model, but rather to evaluate whether the multimodal dataset supports meaningful joint learning across heterogeneous data sources. Such validation is commonly used in Data Descriptors to assess dataset usability and internal consistency. A reference multimodal architecture was implemented to integrate dynamic MRI sequences (T2WI and CE) with structured clinical variables. The workflow is illustrated in Fig. 6. The model was trained using the released dataset and evaluated on an independent test set. Reference multimodal validation framework used for assessing dataset consistency. This figure provides a schematic overview of the reference multimodal prediction pipeline used for technical validation, illustrating the integration of dynamic MRI sequences and structured clinical data for joint regression and classification tasks. The model achieved a mean absolute percentage error (MAPE) of 0.265 for the three treatment-difficulty indicators and an area under the receiver operating characteristic curve (AUC) of 0.931 for the binary recommendation label (Fig. 7). Performance of the reference multimodal validation framework. ROC and PR curves illustrating the classification performance of the reference multimodal prediction pipeline on the independent test set. The model is provided as a baseline implementation for technical validation of the dataset. The 95% confidence intervals and p-values of AUC/AP were estimated using bootstrapresampling with 1,000 iterations. For ROC AUC, the null value was set to 0.5. For AP, the null valuewas set to the positive class prevalence. These results indicate that the imaging and structured data components contain complementary information and can be jointly utilized in multimodal analytical frameworks. Importantly, this validation experiment is provided solely to demonstrate dataset coherence and usability. The reported performance should not be interpreted as representing a clinically validated decision-support system or as a benchmark for future methodological comparisons. Summary of technical validation Collectively, the descriptive statistical analyses, validation of the rule-based labeling strategy,and multimodal consistency validation demonstrate the internal consistency, annotation reliability, and practical usability of the dataset. The dataset captures substantial clinical and procedural heterogeneity and supports a range of analytical approaches, including traditional statistical modeling and modern multimodal machine learning techniques. Usage Notes This dataset is intended for research and educational purposes and is suitable for a range of analytical and methodological applications. Potential use cases include descriptive statistical analysis, development and benchmarking of predictive models, and exploration of multimodal data fusion techniques involving medical imaging and structured clinical data. All MRI images included in the dataset correspond to pretreatment examinations and are provided as continuous sagittal T2-weighted and contrast-enhanced sequences. Users may preprocess the imaging data according to their specific methodological requirements, including normalization, spatial resampling, temporal subsampling, or feature extraction. The FUAS recommendation label is derived from a predefined rule-based decision tree (see Methods) and represents one operational definition of treatment suitability. Users are encouraged to reinterpret, modify, or replace this label according to alternative clinical assumptions or modeling objectives. Similarly, the treatment difficulty indicators—sonication time, EEF, and NPVR—are provided as continuous variables to allow flexible downstream analysis. Users should note that the dataset does not include long-term clinical follow-up information or reproductive outcome data. Consequently, analyses derived from this dataset should be limited to assessments of treatment feasibility, procedural difficulty, and short-term treatment outcomes rather than long-term prognosis. The dataset is not intended to support direct clinical decision-making or patient management without appropriate external validation. To promote reproducibility and transparency, users are encouraged to clearly document all data preprocessing steps, modeling strategies, and evaluation protocols when publishing results based on this dataset, and to cite the associated Data Descriptor. Author contributions Qianru Zeng and Ling Fan contributed equally to this work as co-first authors. Qianru Zeng, Ling Fan, Shuaibin Liu, Min Zhou, and Qian Wang were responsible for data collection and organization. Qianru Zeng constructed and curated the dataset. Yu Zhao performed data preprocessing and conducted model development, implementation, and validation. Qianru Zeng and Ling Fan drafted the manuscript. Li Ling and Min Zhou critically revised the manuscript and approved the final version. Li Ling is the primary corresponding author, and Min Zhou is the co-corresponding author. All authors reviewed and agreed to the submitted version. Funding This work was supported by the National Natural Science Foundation of China (No. 82304959), Chongqing Medical Scientific Research Project (Joint Project of Chongqing Health Commission and Science and Technology Bureau) (No. 2026MSXM136), Project of high-intensity ultrasound focused therapy system promotion and application (No. 2024TGYY35), and Joint Project of Pinnacle Disciplinary Group, the Second Affiliated Hospital of Chongqing Medical University (No. 2025jjcs030). Code availability The source code for data preprocessing, model training, and validation is publicly available via the Figshare repository under the persistent https://doi.org/10.6084/m9.figshare.32508117. The code is provided as a reference implementation for dataset validation and is not intended to represent an optimized or clinically deployable system. Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. These authors contributed equally: Qianru Zeng, Ling Fan. Contributor Information Min Zhou, Email: zhoumin@cqmu.edu.cn. Li Ling, Email: 304869@hospital.cqmu.edu.cn. References Associated Data Data Citations Data Availability Statement The source code for data preprocessing, model training, and validation is publicly available via the Figshare repository under the persistent https://doi.org/10.6084/m9.figshare.32508117. The code is provided as a reference implementation for dataset validation and is not intended to represent an optimized or clinically deployable system.