Optimizing acquisition time and injected dose in 18F-FDG PET/CT imaging using deep learning: enhancing image protocol efficiency and safety Authors: Yulong Zeng, Weixia Chong, Zhao Ge, Jie Qin, Biyun Mo, Xingyu Mu, Wei Fu Published: 2026 ABSTRACT AbstractBackgroundThis study aims to evaluate the effectiveness of deep learning algorithms in simulating standard acquisition time images from shortened acquisition times in 18F-FDG PET/CT imaging, thereby optimizing both image quality and radiopharmaceutical use.MethodsWe evaluated 322 patients who underwent 18F-FDG PET/CT, simulating half-dose conditions and reconstructing images at acquisition times of 30 s, 45 s, 60 s, 75 s, and 90 s while maintaining identical bed positions and anatomical coverage. Images were reconstructed using a residual U-Net architecture with data augmentation to enhance training. Image quality was assessed using a 5-point Likert scale, inter-reader consistency with Cohen’s kappa, and quantitative metrics (PSNR, SSIM, MAE, MSE, RMSE) to compare pre- and post-processing conditions, validating the model’s effectiveness. Lesion detectability was specifically assessed by counting detectable lesions before and after processing, evaluated by two senior nuclear medicine physicians to ensure consistency. Statistical analyses employed Wilcoxon signed-rank tests and kappa statistics, considering p-values less than 0.05 as statistically significant.ResultsDeep learning significantly improved image quality scores in the half-dose group from 3.45 ± 0.48 to 4.33 ± 0.55 and enhanced kappa values from 0.76 to 0.82 (p < 0.05). Quantitative metrics demonstrated marked improvements, with SSIM values rising from 0.75 to 0.87 and PSNR from 37.65 dB to 41.37 dB in half-dose scenarios. Similarly, in the 30s model, SSIM increased from 0.73 to 0.84 and PSNR from 35.38 dB to 40.15 dB. Lesion detectability in small lesions improved by up to 16.3% in shortened acquisition conditions, effectively enhancing the detection of critical pathological features.ConclusionsThe application of deep learning models in PET/CT imaging may improve scan efficiency and reduce radiopharmaceutical usage while maintaining image quality without apparent loss of diagnostic information. These findings support the potential of AI-enhanced imaging protocols to reduce patient radiation exposure and improve clinical workflow efficiency.Keywords: Deep learning, 18F-FDG, PET/CT, Acquisition time, Injection dose Background Positron emission tomography-computed tomography (PET/CT) is a widely used non-invasive imaging technique in oncology that serves as a powerful tool for diagnosis, cancer staging, radiation therapy planning, and treatment response monitoring [1–3]. The PET reconstruction algorithm is crucial for achieving high image quality and precise quantification in imaging diagnostics [4, 5]. In current PET/CT image reconstruction, various conventional post-processing methods such as Gaussian filtering, Metz filtering, wavelet transforms, non-local means, and BM3D techniques are employed to effectively reduce image noise [6, 7]. While these methods are based on robust mathematical principles and are effective in noise suppression, they often compromise the accuracy of lesion detection due to their limited adaptability and the oversimplification of processing parameters [8, 9]. Furthermore, advances in deep learning have significantly enhanced capabilities in image recognition, segmentation, registration, super-resolution, and denoising [10, 11]. For example, Wang et al. developed unsupervised models to reconstruct PET/CT images from sinograms, showing that deep learning methods surpass traditional algorithms like filtered back projection and maximum likelihood expectation maximization in both peak signal-to-noise ratio and structural similarity [12]. Additional studies corroborate these findings, further confirming the role of deep learning in improving image quality across various diagnostic applications [13–17]. Moreover, the practical efficacy of deep learning is evidenced by the FDA’s approval of several algorithms, such as SubtlePET by Subtle Medical, AiCE DLR by Canon Medical Systems, and HYPER DLR by United Imaging, which have enhanced PET/CT image quality and diagnostic accuracy [18, 19]. While previous deep learning approaches have significantly enhanced PET/CT image quality and diagnostic accuracy, they often use conventional architectures that struggle to effectively process subtle medical image features, especially in varied clinical conditions [20, 21]. Recent work by Ciborowski et al. [22] demonstrated the feasibility of applying a U-Net-based deep learning enhancement method to 18F-FDG PET images acquired with reduced acquisition times; however, this approach primarily focused on post-reconstruction image enhancement using standard U-Net architectures. Such limitations in deep feature extraction and a tendency to overfit can restrict their performance on new, unseen data. To bridge these challenges, the residual U-Net architecture integrates residual blocks to prevent vanishing gradients, a common issue in extensive networks. These blocks allow deeper network structures and facilitate the direct flow of gradients across multiple layers, simplifying training and preserving information integrity [23]. Despite its advantages, clinical research into the residual U-Net for medical imaging, particularly in PET/CT, remains sparse [24]. Therefore, our study aims to validate and extend its application to enhance diagnostic precision and robustness across varied clinical conditions. Our study employs the residual U-Net to simulate images from shortened acquisition times and reduced doses, aiming to refine 18F-FDG PET/CT imaging protocols, comparing the effectiveness of these simulations to traditional scans in terms of qualitative image quality, quantitative image metrics, and lesion detectability. This strategy enhances diagnostic precision and reduces patient exposure to radiation, thereby improving medical imaging safety and efficiency. Methods Patients This study involved 322 patients who underwent 18F-FDG PET/CT at Department of Nuclear Medicine of The First Affiliated Hospital of Guilin Medical University from February 2023 to February 2024. Their demographic and clinical data are listed in Table 1. Patients were categorized into normal-weight and overweight groups based on body mass index (BMI), with BMI < 25 kg/m2 defined as normal weight and BMI ≥ 25 kg/m2 defined as overweight. This study adhered to the ethical standards of the Declaration of Helsinki and was approved by the Institutional Review Board of the First Affiliated Hospital of Guilin Medical University (Approval No. 2023-QTLL-34). Due to the retrospective and non-interventional nature of the study, the requirement for written informed consent was waived by the ethics committee. Table 1Patient demographic characteristicsParameterValueAge (years)62.8 ± 11.2aWeight (kg)65.5 ± 15.4aHeight (cm)162.3 ± 10.5aBMI24.8 ± 3.6aInjected activity (Mbq)357.5 ± 45.3aAcquisition time (min)64.0 ± 9.5aDisease Type Lung Cancer70(20.5%)b Liver Cancer45(13.2%)b Prostate Cancer35(10.3%)b Breast Cancer40(11.7%)b Thyroid Cancer25(7.3%)b Stomach Cancer20(5.9%)b Colorectal Cancer30(8.8%)b Lymphoma35(10.3%)b Nasopharyngeal Carcinoma12(3.5%)b Pancreatic Cancer10(2.9%)b Renal Cancer8(2.3%)b Mediastinal Neoplasms7(2.1%)b Melanoma5(1.5%)bBone Sarcoma5(1.5%)bBMI, body mass indexa Data are presented as the mean ± standard deviationb Number of patients and proportions Patient demographic characteristics BMI, body mass index a Data are presented as the mean ± standard deviation b Number of patients and proportions PET/CT imaging All patients were required to fast for over six hours prior to the examination, in accordance with nuclear medicine guidelines. The 18F-FDG PET/CT scans were conducted using two different systems: a Philips Ingenuity TF PET/CT (Philips Healthcare, Eindhoven, The Netherlands), which was used for 244 cases, and a Neuwise NewEra PET/CT scanner (Neusoft Medical Systems Co., Ltd., Shenyang, China), utilized in 78 cases. After synthesizing 18F-FDG in-house via the GE PETtrace 840 cyclotron and China Zhengze X5-FDG synthesis module, patients rested for 60 min. 18F-FDG was administered using a body weight-adjusted activity (3-3.7 Mbq/kg) in accordance with institutional clinical protocols and current guidelines. A preparatory CT scan was then performed on both systems immediately prior to the PET scan, covering the area from the top of the skull to mid-thigh over 10–11 bed positions. Each position was scanned for 90 s at a tube voltage of 120 kV, a tube current of 251 mA, and a slice thickness of 4 mm, following the established protocols recommended by nuclear medicine guidelines. Image reconstruction was performed using a vendor-implemented ordered-subsets expectation maximization (OSEM) algorithm [25]. Low-count PET images were simulated at the list-mode level prior to image reconstruction. Short-acquisition-time images (30 s, 45 s, 60 s, and 75 s) were generated by truncating the acquisition time window of the original standard-protocol list-mode data while maintaining identical bed positions and anatomical coverage. Simulated half-dose images were produced by randomly down-sampling detected coincidence events from the list-mode data before reconstruction to emulate reduced injected activity. This proportional event-based reduction preserves the intrinsic Poisson count statistics of PET acquisition. All reconstruction parameters, including iterations, subsets, time-of-flight modeling, and point-spread function correction, were kept identical across standard and simulated datasets to ensure methodological consistency. Although true reduced-dose acquisitions were not available in this retrospective dataset, list-mode statistical decimation has been clinically validated against real low-dose PET scans in human subjects, demonstrating comparable quantitative performance and noise characteristics [26]. Therefore, this approach is widely accepted as a reliable simulation strategy in low-dose PET research. Figure 1 illustrates representative PET/CT images acquired at different acquisition times and under simulated half-dose conditions. PET/CT images at various acquisition times and half-Dose, All PET images were displayed using identical fixed SUV window settings Model construction Our deep learning framework was developed using image datasets derived from 322 patients acquired under varied acquisition times and simulated half-dose conditions. For each patient, six image sets were generated, including one standard-protocol PET image (90 s, standard injected activity) and five corresponding low-count PET images (30 s, 45 s, 60 s, 75 s, and simulated half-dose), resulting in a total of 322 ✕ 6 = 1932 image sets. All images were resampled to ensure consistent size and resolution. The model was implemented as a 2D residual U-Net operating on axial PET slices and trained using the PyTorch framework, with feature channels of 16, 32, 64, 128, and 256 and four residual units to enhance training depth and efficiency. To diversify the training data, random noise addition, image flipping, rotation, and scaling were applied, improving robustness to imaging variations. Uniform preprocessing and training strategies were applied across all acquisition-time and dose groups. For model training, each low-count image was paired with the corresponding standard-protocol PET image from the same patient as the reference target. Dataset splitting was strictly performed at the patient level using a 3:1:1 ratio, resulting in approximately 194, 64, and 64 patients assigned to the training, validation, and test sets, respectively. All image sets, including all bed positions and axial slices derived from a given patient, were assigned exclusively to a single subset to prevent data leakage. No images from the same patient were distributed across different datasets. Scanner types (Philips Ingenuity TF and Neusoft NewEra) were proportionally represented across the training, validation, and testing cohorts to minimize scanner-specific bias and ensure balanced evaluation. Training was conducted using the Adam optimizer with an initial learning rate of 1 × 10− 5 and an ExponentialLR schedule over 200 epochs with a batch size of 32. During training, model performance was continuously monitored on the validation set, and the model with the lowest validation loss was selected for final evaluation to mitigate overfitting. Employing a residual U-Net architecture, we trained five dedicated models to address half-dose images and reduced acquisition times of 30s, 45s, 60s, and 75s. Five-fold cross-validation was performed within the training cohort to further assess model robustness and stability, while the independent test set was strictly reserved for final performance evaluation. Model training and inference were performed on a workstation equipped with an NVIDIA GeForce RTX 4090 GPU (24 GB memory). Figure 2 illustrates the oerall workflow and network framework. Research workflow and network framework diagram Image quality assessment Image quality was independently assessed by two nuclear medicine physicians, each with over ten years of experience, using a 5-point Likert scale that evaluated lesion contrast and overall image noise. The scoring was as follows: a score of 1 indicated excessive noise and poor lesion contrast; 2 was assigned for sub-optimal noise and blurred lesions affecting diagnostic confidence; 3 denoted equivalence to routine oncological 18F-FDG PET/CT images; 4 suggested preferred noise levels and good lesion contrast with minimal clinical impact from blurring; and 5 represented optimal noise levels and clear lesion contrast [27]. Furthermore, lesion detectability was quantitatively assessed by counting the detectable 18F-FDG-avid lesions in each series across various acquisition times and dose levels, using images from the standard 90s full-dose acquisition as the benchmark. Lesions were classified based on their maximum diameter measured on PET/CT images, with lesions ≤ 10 mm defined as small lesions and those > 10 mm defined as large lesions. Lesion detectability was assessed on a per-patient basis using the predefined index lesion, and paired comparisons were performed between pre- and post-processing images for the same lesion within the same patient. In instances of assessment discrepancies, a third, highly experienced nuclear medicine physician was consulted during a consensus meeting to finalize the evaluations. This analysis was intended as an exploratory assessment rather than a formal evaluation of diagnostic performance. Quantitative image metrics The effectiveness of the deep learning models was rigorously assessed using five quantitative metrics selected for their relevance to image quality and diagnostic accuracy. The peak signal-to-noise ratio (PSNR), defined as , measures the ratio between the maximum possible signal intensity and reconstruction error, thereby reflecting overall image clarity. The structural similarity index (SSIM), calculated as , evaluates perceptual image quality by jointly considering luminance, contrast, and structural information. The mean absolute error (MAE), defined as , quantifies the average absolute difference between the network-generated images and the reference images, providing a direct measure of reconstruction accuracy. The mean squared error (MSE), calculated as , characterizes the average squared deviation and reflects the overall energy of the reconstruction error. Finally, the root mean squared error (RMSE), defined as , describes the dispersion of errors and is particularly sensitive to larger deviations. Together, these metrics provide a comprehensive evaluation of the model’s ability to preserve image quality across different acquisition conditions, supporting the robustness and potential clinical utility of the proposed deep learning approach. Lesion ROIs were defined at the patient level using a single index lesion per patient on the standard 90s reference images. The index lesion was identified based on clinical reports and consensus visual assessment by two experienced nuclear medicine physicians. When multiple lesions were present, the lesion with the highest SUVmax on the reference image was selected as the index lesion. For quantitative analysis, SUVmax values were measured within the index lesion ROI on both the reference and DL-processed images. To assess voxel-level modification patterns, △-maps were generated by subtracting the low-count input image from the DL-processed image. The mean absolute voxel-wise modification magnitude (|△|) was calculated within the lesion ROI. A background ROI was placed in contralateral normal-appearing tissue on the same axial slice, carefully avoiding large vessels and physiologic high-uptake regions. The concentration ratio (R) was defined as the ratio of mean |△| within the lesion ROI to that within the contralateral background ROI. Statistical analysis Statistical analyses were conducted using SPSS 25.0 (IBM, Armonk, NY, USA). Continuous variables were presented as mean ± SD and assessed for normality with the Shapiro-Wilk test. Inter-reader agreement on image quality was evaluated using Cohen’s kappa, and changes in kappa values were statistically assessed using Z-tests to compare pre-processing and post-processing stages. Differences in image quality scores and lesion detectability across various groups and conditions were analyzed using the Wilcoxon signed-rank test for paired comparisons and paired t-tests or non-parametric tests as appropriate, compared between independent groups using the Mann-Whitney U test. A p-value of less than 0.05 was considered statistically significant. Results Qualitative image quality Our analysis confirmed significant enhancements in image quality and inter-reader consistency across various groups after the application of our deep learning techniques. Notably, in the half-dose group, image quality scores significantly increased from a pre-processing mean of 3.45 ± 0.48 to a post-processing mean of 4.33 ± 0.55, demonstrating the effectiveness of our deep learning model under reduced radiopharmaceutical conditions (p < 0.05). Similarly, the 30s and 45s acquisition groups displayed marked improvements; scores rose from 3.10 ± 0.69 to 3.89 ± 0.65 and from 3.43 ± 0.52 to 4.12 ± 0.52, respectively, both with p-values below 0.05. The kappa values, indicative of inter-reader agreement, also demonstrated significant improvement, suggesting enhanced consistency in image interpretation between physicians. For instance, in the half-dose group, kappa values increased from 0.76 to 0.82 (p < 0.05). The 30s group showed an increase from 0.68 to 0.75, and the 45s group from 0.75 to 0.81, both with p-values less than 0.05. These increases affirm the correlation between enhanced image quality and improved diagnostic consistency. Table 2 presents a comparative assessment of pre- and post-processing image quality scores and kappa values across all acquisition groups. Statistically significant improvements were observed primarily in the half-dose, 30 s, and 45 s groups, while the 60 s and 75 s groups demonstrated more modest changes. Table 2Comparative assessment of image quality and inter-reader agreement across groupsGroupMetricPre-Processing ScorePost-Processing Scorep-valueHalf-doseImage Quality Scores3.45 ± 0.484.33 ± 0.55< 0.05 30s3.10 ± 0.693.89 ± 0.65< 0.05 45s3.43 ± 0.524.12 ± 0.52< 0.05 60s3.56 ± 0.394.20 ± 0.48< 0.05 75s3.94 ± 0.374.63 ± 0.510.32Half-doseKappa Values0.760.82< 0.05 30s0.680.75< 0.05 45s0.750.81< 0.05 60s0.790.83< 0.05 75s0.820.840.25 Comparative assessment of image quality and inter-reader agreement across groups Quantitative image metrics Our analysis suggests that deep learning models can maintain image quality across various dosage conditions. Specifically, in half-dose scenarios, the model maintained high image quality with an SSIM improving from 0.75 to 0.87 and a PSNR increasing from 37.65 dB to 41.37 dB, metrics that, although slightly below the maximum observed, still exceed clinical thresholds. The RMSE measured at 2.82 indicates a moderate increase in prediction errors compared to longer and full-dose acquisition times, yet within acceptable limits, illustrating the model’s robustness under reduced dose conditions. Regarding the model performance developed from images acquired over shorter durations, SSIM was improved from 0.73 to 0.84 for the 30s model and from 0.76 to 0.85 for the 45s model, showcasing the model’s ability to preserve image quality effectively. Similarly, the PSNR for the 30s model improved from 35.38 dB to 40.15 dB, and for the 45s model from 37.03 dB to 41.73 dB, suggesting that image quality remains within clinically acceptable ranges even when the model is trained with significantly reduced acquisition time data. For the 75s model, we observed a notable improvement in image quality, with SSIM increasing from 0.87 to 0.93 and PSNR rising from 42.88 dB to 44.31 dB, indicating the efficacy of deep learning even at near-standard acquisition times. However, the modest enhancement reflects the limited scope for significant improvement, given the already high baseline quality. These results underscore the deep learning model’s capability to adapt to and compensate for shorter imaging times without sacrificing diagnostic integrity. Table 3 provides detailed quantitative assessments for each group. Lesion-based quantitative analysis demonstrated that SUVmax values of DL-processed images were in close agreement with those of the 90 s reference images across all acquisition settings, with no significant systematic differences observed (all p > 0.05; Table 4). Representative examples of the visual improvement achieved by the deep learning model are shown in Fig. 3, comparing the original 30 s acquisition image, the DL-processed image, and the 90 s reference image. Table 3Quantitative image metrics by groupGroupMAEMSERMSESSIMp-valuePSNRp-valueHalf-dose1.217.952.82pre0.75<0.0537.65<0.05post0.8741.37 30s1.4510.563.25pre0.73<0.0535.38<0.05post0.8440.15 45s1.379.543.09pre0.76<0.0537.03<0.05post0.8541.73 60s1.157.232.69pre0.79<0.0539.13<0.05post0.8842.90 75s0.923.571.89pre0.870.7842.880.17post0.9344.31 Quantitative image metrics by group Table 4Lesion SUV agreement with 90s reference in test cohortGroupSUVmax (90 s Ref)SUVmax (Pre)SUVmax (DL Post)ΔSUVmax (Post–Pre)ΔSUVmax (Post–Ref)P (Post–Pre)P (Post–Ref)Half-dose8.52 ± 3.208.71 ± 3.388.44 ± 3.15−0.27 ± 0.56-0.08 ± 0.420.010.37 30 s8.52 ± 3.208.88 ± 3.458.60 ± 3.28−0.28 ± 0.60+ 0.08 ± 0.550.020.29 45 s8.52 ± 3.208.69 ± 3.308.48 ± 3.18−0.21 ± 0.44-0.04 ± 0.380.030.46 60 s8.52 ± 3.208.60 ± 3.228.50 ± 3.19−0.10 ± 0.36-0.02 ± 0.300.080.61 75 s8.52 ± 3.208.56 ± 3.218.53 ± 3.22−0.03 ± 0.22+ 0.01 ± 0.250.210.78 Lesion SUV agreement with 90s reference in test cohort Three representative image sets are shown: the original 30s acquisition image, the DL-processed image, and the 90s reference image Additionally, our study further analyzed the impact of deep learning enhancements on image quality, stratified by patient body weight categories—normal weight and overweight. The enhancements in PSNR were notably more significant in overweight patients. For instance, in the half-dose scenarios, overweight patients exhibited an increase in PSNR from 38.8 dB to 41.7 dB. This trend likely reflects the higher baseline decrease in image quality experienced by overweight patients due to more significant attenuation effects, which our deep learning models effectively compensated for. The model’s capacity to significantly enhance PSNR in these patients underscores its potential to address disparities in image quality aggravated by higher tissue density. Table 5 summarizes these results, providing a clear comparison of PSNR improvements across various patient weights and acquisition times, reinforcing the model’s effectiveness across diverse clinical scenarios. Table 5Detailed PSNR values pre- and post-processing by body weight and acquisition timeGroupWeightPre-Processing PSNR(dB)Post-Processing PSNR(dB)Percentage ImprovementHalf-doseNormal Weight42.144.1+ 4.75%Overweight38.841.7+ 7.47%30sNormal Weight41.843.5+ 4.07%Overweight38.241.0+ 7.33%45sNormal Weight42.344.0+ 4.02%Overweight38.741.5+ 7.24%60sNormal Weight42.744.4+ 3.98%Overweight39.041.8+ 7.18%75sNormal Weight43.544.7+ 2.76%Overweight41.442.1+ 1.96% Detailed PSNR values pre- and post-processing by body weight and acquisition time To preliminarily assess potential scanner-related variability, a subgroup analysis was performed according to the two PET/CT systems used in this study. Quantitative image metrics of DL-processed images showed comparable performance between the Philips Ingenuity TF and Neusoft NewEra scanners, with SSIM values of 0.88 ± 0.03 and 0.87 ± 0.04 and PSNR values of 42.1 ± 2.3 dB and 41.8 ± 2.5 dB, respectively (all p > 0.05). In addition, SUVmax measurements of DL-processed images remained in close agreement with the reference images for both scanners, with mean ΔSUVmax(Post–Ref) values of − 0.04 ± 0.41 and + 0.02 ± 0.45, respectively. These findings suggest that the proposed deep learning model maintains stable quantitative performance across different PET/CT systems. Lesion detectability Our analysis of lesion detectability underscores substantial enhancements achieved through deep learning, particularly for small lesions under varied acquisition times. Lesion detectability was quantitatively assessed by counting detectable lesions before and after deep learning processing, with evaluations independently conducted by two senior nuclear medicine physicians to ensure accuracy and consistency in identifying lesions across varying image conditions. Notably, small lesion detection rates experienced significant improvements in the 30s, 45s, and 60s groups by 16.3%, 13.7%, and 12.9% respectively. These increases are particularly important for early diagnosis, where detecting subtle pathological features is crucial for effective treatment planning.An illustrative clinical example demonstrating the enhancement of small lesion visibility after deep learning processing is presented in Fig. 4. The image demonstrates the effectiveness of our deep learning model in enhancing PET/CT imaging. Prior to processing, the image exhibits high levels of noise, making it difficult to distinguish lung nodules smaller than 1 cm from background noise, which could potentially compromise diagnostic accuracy. After processing, there is a noticeable improvement in image quality, with small lesions becoming distinctly identifiable Conversely, large lesion detectability showed more modest improvements, such as in the half-dose group, where rates marginally rose from 95.3% to 96.2%. This minimal increase suggests that the impact of acquisition condition reductions is less pronounced on larger lesions, which typically have higher initial detectability. The comprehensive enhancements across different patient groups and imaging conditions, as detailed in Table 6, suggest that the deep learning model may improve lesion detectability under challenging imaging conditions. This adaptability highlights its potential to enhance clinical outcomes by providing more reliable and nuanced imaging assessments. Table 6Lesion detection rates pre- and post-processingGroupLesion SizePre-Processing Detection RatePost-Processing Detection RatePercentage ImprovementHalf-doseLarge95.3%96.2%+ 0.9%Small71.4%85.6%+ 13.8% 30sLarge92.5%94.3%+ 1.8%Small65.5%81.8%+ 16.3% 45sLarge94.7%95.3%+ 0.6%Small69.8%83.5%+ 13.7% 60sLarge95.4%96.1%+ 0.7%Small74.3%87.2%+ 12.9% 75sLarge97.1%97.4%+ 0.3%Small80.4%90.1%+ 9.7% Lesion detection rates pre- and post-processing To further explore the subgroup differences observed in lesion size and body weight, voxel-wise △-map analysis was performed on the held-out test cohort. The mean absolute voxel-level modification magnitude (|△|) within the index lesion ROI was significantly higher in small-lesion cases compared with large-lesion cases (p < 0.05). Similarly, overweight patients demonstrated greater |△| values within lesion ROIs than normal-weight patients (p < 0.05). These findings indicate that the network introduced stronger voxel-level intensity adjustments under more challenging imaging conditions, including small lesion size and increased body habitus (Table 7). These findings should be interpreted as exploratory and hypothesis-generating rather than as a comprehensive evaluation of diagnostic performance. Table 7△-map ROI concentration ratio comparisonGroupSmall Lesion R (n = 35)Large Lesion R (n = 29)P valueOverweight R (n = 19)Normal Weight R (n = 45)P valueHalf-dose1.52 ± 0.311.21 ± 0.240.0011.60 ± 0.341.28 ± 0.260.00330 s1.44 ± 0.291.15 ± 0.220.0031.49 ± 0.301.22 ± 0.230.00845 s1.36 ± 0.251.12 ± 0.200.0101.40 ± 0.271.18 ± 0.210.01560 s1.30 ± 0.221.09 ± 0.180.0181.32 ± 0.241.15 ± 0.190.02175 s1.24 ± 0.201.05 ± 0.160.0371.27 ± 0.221.12 ± 0.180.042 △-map ROI concentration ratio comparison Discussion Our analysis demonstrated that deep learning processing notably enhanced image quality, evidenced by the increase from pre-processing scores of 3.45 ± 0.48 to post-processing scores of 4.33 ± 0.55 in the half-dose group, with comparable improvements observed in the 30s and 45s groups. These results suggest the efficacy of our deep learning model in enhancing image quality under reduced radiopharmaceutical conditions. Although SSIM decreased from 0.93 to 0.84 and PSNR from 44.31 dB to 40.15 dB as acquisition durations shortened from 75 s to 30 s, both metrics remained within ranges comparable to standard-protocol images. The relatively small improvement observed in the 75s group, with image quality scores increasing from 3.94 ± 0.37 to 4.63 ± 0.51, likely reflects the minimal reduction in acquisition time compared with the standard 90s protocol. Consequently, the inherent decline in image quality for the 75s scans is less pronounced, resulting in only moderate differences between the pre- and post-processing image quality scores. In addition, lesion-based quantitative analysis demonstrated that SUVmax derived from DL-processed images remained highly consistent with those measured on the 90s reference images, indicating that the proposed deep learning enhancement preserved quantitative consistency while improving image quality. Additionally, the inclusion of data from two distinct PET/CT scanners enhances the external validity of our model by mitigating machine-specific biases and demonstrating its adaptability across varied clinical environments. Consistent with this observation, subgroup analysis across the two scanners demonstrated comparable quantitative performance, supporting the model’s broader applicability in real-world settings. In our study, the kappa values, which measure inter-reader agreement, exhibited significant improvements across various groups, reinforcing the correlation between image quality and diagnostic consistency. For instance, in the half-dose group, kappa values rose from 0.76 to 0.82, and similar increases were observed in the 30s and 45s groups, with kappa values climbing from 0.68 to 0.75 and from 0.75 to 0.81, respectively, both achieving statistical significance (p < 0.05). These enhancements in kappa values suggest that as image quality improves, the likelihood of discrepancies in lesion interpretation among different physicians diminishes, leading to more uniform diagnostic assessments. These findings indicate improved inter-reader consistency following deep learning processing, which may contribute to more consistent image interpretation. Our subgroup analysis of lesion size reveals that reductions in acquisition parameters lead to detectability decreases across all lesions, with smaller lesions particularly affected. This observation highlights a prevalent challenge in PET/CT imaging: enhancing the detection of small, low-contrast lesions under limited conditions. The significantly greater improvements in detection rates for small lesions are partly due to their lower initial detection rates, illustrating the model’s effectiveness in addressing this specific challenge. Regarding body weight, our model markedly enhances image quality for overweight patients, who often suffer from poorer baseline image quality due to increased tissue density. This improvement reduces diagnostic disparities between different weight groups and highlights the model’s ability to adapt to patient-specific challenges. The greater benefit observed in smaller lesions may be related to their higher susceptibility to noise, contrast degradation, and partial volume effects under low-count conditions, making them more responsive to denoising and contrast-enhancing effects of deep learning. Similarly, in larger patients, increased photon attenuation and scatter typically lead to lower signal-to-noise ratios, which may provide greater opportunity for performance gains through AI-based enhancement. Although deep learning models are data-driven and do not explicitly encode physical models such as time-of-flight or point spread function corrections, the observed trends are consistent with known imaging physics principles. To further examine this observation, Δ-map–based analysis was performed on the held-out test cohort. The mean absolute voxel-level modification magnitude (|△|) within the index lesion ROI was significantly higher in small-lesion cases compared with large-lesion cases (p < 0.05). This finding indicates that the network introduced stronger intensity adjustments in small-lesion scenarios, which may partly contribute to the observed improvement in detectability. Moving forward, our research will concentrate on further optimizing the model’s performance for overweight patients and enhancing small lesion detection. This targeted approach aims to advance the model’s capabilities in these challenging aspects, increasing the utility of medical imaging across diverse patient scenarios. Although both reduced injected dose and shortened acquisition time lead to lower detected count statistics, they were not assumed to be physically equivalent in this study. Instead, these two conditions were treated as independent acquisition scenarios and simulated separately using list-mode data. Recent research in PET/CT imaging has increasingly focused on leveraging deep learning technologies to enhance image quality while reducing both radiopharmaceutical dosage and scan duration [28–32]. This shift aims to reduce patient radiation exposure while maintaining diagnostic accuracy in a clinical setting. For example, in one study [33], 61 patients were prospectively enrolled to undergo PET/CT scans with a two-thirds standard dose, followed by a comparison between images reconstructed using deep learning algorithms and those from standard doses. The results indicated that AI-enhanced, dose-reduced images exhibited comparable image quality to standard dose images. Another study [34] utilized images reconstructed with block sequential regularized expectation maximization (BSREM) as training targets, evaluating the deep learning model’s capability to handle reduced acquisition times and lower injection doses-both decreased to 50% of the standard. Yan et al. [35] included 52 patients in their study, using images reconstructed with the OSEM method as training targets. Their findings showed significant improvements in images with acquisition times reduced to 33.3% of the standard, while images with acquisition times shortened to between 25% and 50% maintained lesion detectability comparable to standard acquisition times. Additionally, this research included a subgroup analysis based on patient weight, revealing that overweight and obese patients could have their FDG dosage reduced by up to 50% compared to the standard. The integration of advanced deep learning frameworks, notably the residual U-Net architecture [36–38], substantially enhances our study’s outcomes. Residual U-Net evolves from traditional U-Net by incorporating residual blocks that facilitate training of deeper networks. These blocks enable gradient flow through alternative pathways, enhancing the network’s capability to discern complex patterns and fine details in datasets typical of scenarios with reduced doses and shortened acquisition times [39, 40]. Residual connections mitigate the vanishing gradient issue common in deep neural networks, promoting effective error backpropagation [41]. By refining the detection of nuanced features often lost in deeper network layers, the residual U-Net demonstrates its ability to reconstruct high-quality images from sparse input data [42], including those from half-doses or shortened scans, while preserving quantitative similarity and lesion detectability metrics, aligning with the current trends in patient-centered medical imaging. Our study presents several innovations: first, the application of the residual U-Net architecture, which significantly enhances image reconstruction by allowing deeper network training and effective error backpropagation, facilitating the detection of intricate patterns and fine details. Secondly, we conducted detailed subgroup analyses on lesion size and patient body weight, demonstrating the model’s effectiveness in enhancing detectability and image quality under varied clinical conditions, particularly for small lesions and overweight patients. Lastly, our model effectively addresses the dual challenges of reducing radiopharmaceutical dosage and shortening acquisition times, demonstrating substantial advancements in reducing radiation exposure and improving operational efficiency. A limitation of our study is the exclusive use of a single radiotracer, 18F-FDG, which, despite its wide use in oncology, may not fully represent the metabolic or molecular activities of all cancer types or pathological conditions. This limitation could restrict the generalizability of our findings, especially where 18F-FDG is less effective. Additionally, using only two PET/CT scanners could introduce machine-specific biases, potentially limiting the wider applicability of our results. Furthermore, the present study did not include direct comparisons with conventional post-reconstruction denoising filters or other AI-based methods, which may be explored in future work. To address these limitations, future studies should employ a broader range of radiotracers to cover various cellular processes and disease markers, thereby enhancing the training and validation of deep learning models across diverse imaging scenarios. Expanding the use of different PET/CT systems may further reduce technology-specific biases and improve generalizability. Although the present work was conducted in the image domain using reconstructed PET images, future investigations may explore projection- or sinogram-domain deep learning approaches to more directly integrate acquisition physics and potentially enhance reconstruction performance. Additionally, developing models tailored to different lesion sizes and patient weights may further improve diagnostic performance in heterogeneous patient populations. Multi-modal imaging techniques integrating PET with CT, MRI, or ultrasound could provide more comprehensive insights into disease characterization and treatment response [43–46], further supported by advances in novel probes and AI-driven personalized imaging strategies. Conclusions Our findings suggest that deep learning can effectively optimize PET/CT imaging by lowering both radiopharmaceutical doses and acquisition times. This approach has the potential to lower patient radiation exposure and improve patient comfort and clinical workflow efficiency. Additionally, this model processing preserved quantitative similarity and improved lesion detectability, with potential implications for clinical workflow optimization. Acknowledgements Not applicable. Abbreviations 2-deoxy-2-[18F]fluoro-D-glucose Artificial intelligence Block-matching and 3D filtering Body mass index Computed tomography Deep learning Food and Drug Administration Graphics processing unit Institutional Review Board Mean absolute error Megabecquerel Magnetic resonance imaging Mean squared error Ordered-subsets expectation maximization Positron emission tomography Positron emission tomography-computed tomography Point-spread function Peak signal-to-noise ratio Region of interest Root mean squared error Standard deviation Statistical Package for the Social Sciences Structural similarity index Standardized uptake value Maximum standardized uptake value Time of flight U-shaped convolutional neural network Author contributions Zeng Yulong designed the study, performed the data analysis, and wrote the manuscript. Ge Zhao conducted the data analysis. Chong Weixia and Mo Biyun collected the data. Qin Jie contributed to both the data collection and analysis. MuXingYu and Fu Wei participated in the study design and manuscript revisions. All authors read and approved the final manuscript. Funding This work was supported by the Natural Science Foundation of Guangxi Zhuang Autonomous Region (2025GXNSFHA069001), the National Natural Science Foundation of China (82402332; 82460356) and Guangxi Health and Family Commission (Z-C20220815). Data availability The datasets generated and/or analyzed during the current study are not publicly available due to privacy and confidentiality concerns, as well as legal and ethical restrictions. However, the data are available from the corresponding author on reasonable request. Declarations Ethics approval and consent to participate This retrospective study was conducted in accordance with the Declaration of Helsinki and was approved by the Institutional Review Board of The First Affiliated Hospital of Guilin Medical University (Approval No. 2023-QTLL-34). The requirement for written informed consent was waived by the ethics committee because of the retrospective and non-interventional nature of the study. Consent for publication Not applicable. 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. Xingyu Mu and Wei Fu contributed equally to this work. Contributor Information Xingyu Mu, Email: muxingyu1992@hotmail.com. Wei Fu, Email: fuwei19700513@163.com. References Associated Data Data Availability Statement The datasets generated and/or analyzed during the current study are not publicly available due to privacy and confidentiality concerns, as well as legal and ethical restrictions. However, the data are available from the corresponding author on reasonable request.