两款深度学习模型用于经会阴超声视频自动检测压力性尿失禁的回顾性研究
Deep learning–based detection of stress urinary incontinence in transperineal ultrasound video examinations
在德国亚琛大学医院一项纳入803例患者的回顾性单中心研究中,USCL+TF与V-JEPA2两种深度学习方法均可对完整矢状面经会阴超声(TPUS)视频进行分类,无需手动选帧或定义感兴趣区。
提供两款深度学习方法在无手动选框下完成压力性尿失禁TPUS视频分类的性能对比及可行性结论。
Abstract
Stress urinary incontinence (SUI) is a common bladder dysfunction in women, and transperineal ultrasound (TPUS) enables dynamic assessment of the bladder neck, urethra, and pelvic floor. We developed and compared two deep-learning approaches for automated SUI detection from complete midsagittal TPUS videos with adequate urethral visualization and without manual frame or region-of-interest selection. In this retrospective single-center study, 803 TPUS recordings from unique patients were analyzed. Two video-classification pipelines were evaluated: an ultrasound-pretrained frame-wise encoder with transformer-based temporal aggregation (USCL + TF) and a video-native foundation model (V-JEPA2). For each architecture, we tested a standard input configuration and two anatomically guided variants based on automated segmentation: one restricting the input to an automatically defined region of interest and one augmenting the input with segmentation masks. Both model families showed good discriminative performance. Across 50 repeated paired splits, USCL + TF achieved mean area under the receiver operating characteristic curve (AUC) values from 0.84 to 0.88, whereas V-JEPA2 achieved AUC values from 0.89 to 0.90. In paired analyses accounting for overlapping splits and seven comparisons, V-JEPA2 had a higher AUC than USCL + TF in the basic configuration. The other matched architecture comparisons and the comparisons of anatomy-guided with basic inputs were not statistically significant after multiplicity correction. These findings show that deep learning can classify SUI from full TPUS video examinations with good performance and without manual frame selection, supporting further validation of automated TPUS analysis as a decision-support tool.
Introduction
Stress urinary incontinence (SUI) is one of the most prevalent pelvic floor disorders in women, affecting millions worldwide and substantially impairing quality of life1,2,3. SUI is defined as the involuntary leakage of urine during activities that increase intra-abdominal pressure, such as coughing, sneezing, or physical exertion4. The two principal pathophysiological mechanisms are urethral hypermobility and intrinsic sphincter deficiency, which are often associated with pelvic floor weakening after childbirth and during the menopausal transition5,6,7. The traditional diagnostic workup includes medical history and clinical examination and, in selected cases, urodynamic testing5.
Over recent decades, transperineal ultrasound (TPUS) has become an important modality in urogynecologic imaging8. TPUS is a cost-effective, noninvasive technique that enables dynamic, real-time assessment of the urethra, bladder neck, and pelvic floor structures during maneuvers such as Valsalva and coughing9,10,11. It can be used to derive measurements such as bladder neck motion and urethral rotation, which are directly relevant to SUI assessment and can complement clinical examination and urodynamic testing12,13. However, image acquisition and quantitative assessment remain highly dependent on operator experience and may show considerable inter-observer variability, which limits reproducibility and standardization14,15,16. Moreover, dynamic interactions among periurethral structures are complex, and currently used measurements may not fully capture all relevant functional abnormalities, potentially resulting in incomplete diagnostic assessment.
To address these limitations, artificial intelligence (AI), particularly deep learning, may provide a more scalable and standardized approach. In pelvic floor imaging, deep learning methods have shown promising performance in segmentation, anatomic measurement, and classification tasks17,18,19. For SUI specifically, previous studies have also reported encouraging diagnostic accuracy. However, many workflows still rely on manual frame selection or manually defined regions of interest20,21.
To reduce manual input preparation and move toward workflow integration, we developed and compared two deep-learning strategies for SUI detection using complete midsagittal TPUS videos. For each examination, the central sagittal plane was extracted automatically from the 4D volume, after which a gynecologist confirmed adequate urethral visualization in the resulting video. After this plane extraction and quality check, neither strategy required manually selected diagnostic frames or regions of interest. One model leverages representations from pretraining on clinical ultrasound images, whereas the other leverages representations from natural-video data. We also evaluated domain knowledge-guided input configurations to assess whether they improve performance. Finally, we performed explainability analyses to examine whether the learned decision patterns were qualitatively aligned with clinical expectations.
The objective of this study was to establish an end-to-end video-based framework for SUI detection that avoids manual frame or region-of-interest selection while maintaining strong diagnostic performance. By directly comparing two architecture families and multiple input strategies in a clinically acquired TPUS cohort, this study aims to inform practical design choices for AI-assisted SUI assessment and to support future validation toward clinical integration.
Methods
Study population
This retrospective study was conducted following local data protection regulations. After approval by the local Ethical Committee (Medical Faculty, RWTH Aachen University (Germany): EK EK028-19), the requirement for individual informed consent was waived.
The study included patients seen in the urogynecology consultation at the Incontinence Center of University Hospital Aachen between January 2014 and December 2023. Patients with suspected urinary incontinence or genital prolapse were referred for a standardized diagnostic workup that included the validated ICIQ-UI Short Form questionnaire, a detailed urogynecological history, urinalysis, and vaginal examination. The vaginal examination was preferably performed with a comfortably full bladder and included assessment for pelvic organ prolapse. A standardized cough stress test was also performed, with repositioning of prolapse when necessary to evaluate demonstrable stress urinary leakage. When suspected SUI required further clarification, patients underwent urodynamic testing comprising uroflowmetry, cystometry, and urethral pressure measurements. TPUS complemented this workup by assessing pelvic floor anatomy and excluding relevant paraurethral pathology. The physician’s diagnosis of SUI was based on these combined clinical findings. All examinations and diagnoses were performed by board-certified gynecologists with 6–16 years of clinical experience (increasing over the study period) and certified according to DEGUM level II and AGUP level III standards.
The initial imaging collection comprised 9,893 ultrasound examinations, predominantly from the urogynecology consultations described above. After 7,768 repeat visits and follow-up examinations were excluded to limit the inclusion of postoperative examinations, 2,125 first-visit examinations remained. Each examination could contain multiple ultrasound recordings. Preview sagittal center still frames were screened manually to identify TPUS recordings. Non-TPUS recordings, including perianal ultrasound and other examination types, were excluded. If an examination contained more than one TPUS recording, the first TPUS recording was retained. This yielded one TPUS recording from each of 1,295 examinations, while 830 examinations without a TPUS recording were excluded. The corresponding urogynecological consultation reports were then retrieved from the hospital information system. A usable report containing a final diagnosis was unavailable for 355 examinations, leaving 940 examinations with one TPUS recording and an available diagnosis. The 940 retained TPUS recordings were converted automatically to midsagittal videos. These videos were assessed by a gynecologist with 18 years of experience who was blinded to the reference diagnosis. A total of 137 recordings were excluded because the urethra was not sufficiently visible in the automatically selected plane, leaving 803 recordings from 803 patients (Fig. 1). No additional explicit inclusion or exclusion criteria were applied.
For this study, the label was taken from the final diagnosis documented in the corresponding urogynecological consultation report. Examinations in which SUI was documented, including mixed urinary incontinence with a stress component, were labeled SUI-positive, while examinations without a documented SUI diagnosis were labeled SUI-negative. The SUI-negative group could include urge incontinence and other urogynecological diagnoses. The model therefore predicts the clinically established SUI diagnosis from TPUS recordings alone. The clinical diagnosis was not independently readjudicated for this study.
A randomly selected subset of 95 recordings was used for manual segmentation annotation. This subset was assigned exclusively to the training set for the classification tasks. In each repetition, the remaining 708 recordings were allocated by stratified random sampling to 388 training, 160 validation, and 160 test examinations, resulting in overall partition sizes of 483, 160, and 160, respectively, while preserving an SUI prevalence of approximately 42% across splits.
Cohort selection and dataset partitioning. Exclusion counts indicate the number removed at each step.
TPUS acquisition and preprocessing
TPUS examinations were performed with the patient in the lithotomy position and with a comfortably full urinary bladder. The transducer was placed on the perineum and oriented approximately 45° ventrally toward the symphysis pubis. The acquisition protocol aimed to maintain a standardized probe position and anatomical orientation, with visualization of the symphysis pubis, bladder neck, bladder base, and relevant portions of the levator ani and pelvic floor. The standardized sequence consisted of rest, maximal pelvic-floor contraction followed by relaxation, maximal Valsalva, and several forceful coughs. Data were acquired using Voluson E8, S10, and E10 systems (GE Medical Systems).
The raw data consisted of volumetric videos. To reduce computational complexity for downstream analysis, the midsagittal plane was extracted, which typically shows the lower abdomen with the bladder and urethra in view. This plane was selected because it captures urethral motion and bladder dynamics with high diagnostic relevance while requiring substantially less computation than full 3D volumetric video processing.
For model training, videos were preprocessed by center-cropping to a square aspect ratio, resizing to 256 × 256 pixels, and normalizing intensities using ImageNet normalization22. The videos contained a mean of 50.4 ± 18.7 frames at 3.5 ± 1.2 frames/s and had a mean duration of 14.3 ± 2.7 s. A fixed sequence length of 64 frames was set. Shorter sequences were zero-padded with black frames, and longer sequences were downsampled by linear temporal sampling.
Model development
Two deep-learning architecture families were trained and evaluated for video-based classification of stress urinary incontinence: USCL + TF and V-JEPA2. For each architecture, three input configurations were investigated: basic, crop, and crop+append. An overview of the study design and model configurations is shown in Fig. 2.
Overview of model training, including the three input configurations (basic, crop, and crop+append), segmentation-assisted input generation, and parallel evaluation of the USCL + TF and V-JEPA2 spatiotemporal deep-learning models. Segmentation overlays show the bladder in yellow, urethra in red, and symphysis pubis in green. The white outline indicates the automatically derived crop box.
Input configurations
-
1.
In the basic configuration, models received the standard TPUS video input without additional preprocessing beyond the previously described preparation steps.
-
2.
In the crop configuration, each video was spatially restricted to a region defined by the available predicted urethra and symphysis pubis masks throughout the entire sequence. The segmentation model processed every frame automatically. The minimum and maximum coordinates of the predicted urethra and symphysis pubis masks were combined across all frames to define one global bounding box for the video. If no predicted mask was available for crop construction, the full frame would be used as a fallback. If the resulting crop measured less than 100 pixels in height or width, it was expanded to a minimum size of 100 × 100 pixels. The cropped region was then zero-padded to obtain a square input and resized back to 256 × 256 pixels. The aim of this configuration was to reduce irrelevant surrounding context and focus the model on anatomically relevant structures.
-
3.
In the crop+append configuration, the same cropped video input was used, but segmentation information was additionally provided as input. To enable this, the first convolutional layer of each model was replaced with a newly initialized layer accepting six input channels instead of three. These six channels comprised the original three RGB ultrasound image channels and three one-hot encoded mask channels representing the urethra, bladder, and symphysis pubis.
Segmentation-assisted input generation
The crop and crop+append configurations required automated segmentation of relevant pelvic structures. For this purpose, a 3D nnU-Net model23 was trained to segment the urethra, bladder, and symphysis pubis in TPUS videos. A random subset of 95 videos was manually annotated by a medical student after structured training. The subset size was based on the resources available for manual annotation and our prior experience with segmentation model development. Segmentation performance was evaluated using five-fold cross-validation on this annotated subset, yielding a Dice score of 0.79 ± 0.01 (mean ± SD across folds). Mean Dice scores for the bladder, urethra, and symphysis pubis were 0.85, 0.82, and 0.70, respectively. These results were comparable to previously reported results17, but only describe agreement with the held-out available annotations. The trained segmentation model was then applied to all 803 videos to generate the masks used for cropping and mask-appended input generation. After that, an automated check assessed segmentation availability and crop generation.
USCL + TF
The first architecture follows a two-stage strategy in which spatial features are extracted frame-wise using a pretrained 2D network, and temporal dependencies are modeled by a lightweight transformer24,25. USCL26 was selected as the 2D backbone because it is contrastively pretrained on ultrasound video and has been reported to outperform the same ResNet27 architecture with standard ImageNet22 pretraining on downstream ultrasound tasks. In this setup, each frame is processed independently by USCL to produce one feature vector per frame. Learnable positional embeddings are added, and the resulting feature sequence is passed through a shallow three-layer transformer. The final transformer output is then passed to a linear classification head. The model is trained end-to-end with the 2D backbone unfrozen.
V-JEPA2
The second approach fine-tunes V-JEPA228, a native video model pretrained on approximately 1.7 million hours of internet video and ImageNet22. Ultrasound videos were not an explicit component of this pretraining, as the model was designed as a general-purpose video backbone for downstream tasks such as video question answering and robotic planning.
Because V-JEPA2 is designed for sequences of up to 64 frames, it can be directly applied to the previously described input format. An attentive classification probe is added, consistent with the downstream task design described in the original paper. In contrast to many downstream experiments, the video encoder was not frozen in this setting because the visual domain shift from natural videos to ultrasound is substantial.
Training details and code availability
Both architectures shared the same data pipeline: videos were resized to 256 × 256 pixels and sampled to 64 frames. Binary cross-entropy with logits was used as the loss function, and models were optimized with Adam. USCL + TF was trained for 20 epochs with a batch size of 16 and a learning rate of 1 × 10⁻⁴. V-JEPA2 was trained for 10 epochs with a batch size of 2 and a learning rate of 1 × 10⁻⁵. The model checkpoint with the highest validation AUC was retained for evaluation on the test set. Training and evaluation were repeated across 50 stratified random splits. The same training, validation, and test partition were used for all six model configurations within each repetition, allowing paired comparisons. All models were trained on a single RTX 4090 GPU using PyTorch with fp16 mixed precision. We make model and training code available under: https://github.com/spacecheck/sui-tpus.
Statistical analysis
For each model architecture and input configuration pair, classification performance was evaluated on the test set across 50 classification training repetitions using different stratified random splits. Each test set contained 160 patients. Final performance was summarized using mean values and empirical 2.5th − 97.5th percentiles across repetitions for the area under the receiver operating characteristic curve (AUC), sensitivity, specificity, and accuracy. These percentile ranges describe variability across repeated splits of this cohort under the specified splitting procedure. The classification threshold was determined by maximizing Youden’s J statistic on the validation predictions and was then applied unchanged to the corresponding test predictions. AUC was used as the primary comparative measure.
Seven paired AUC comparisons were defined: V-JEPA2 versus USCL + TF within each of the three input configurations, and crop versus basic and crop+append versus basic within each architecture. Mean within-split AUC differences were evaluated using the corrected resampled t test29, which accounts for dependence caused by overlap among repeated splits. For the 50 within-split differences, the corrected standard error was calculated as \(\:SE=\sqrt{\left(\frac{1}{50}+\frac{160}{483}\right){s}^{2}}\), where \(\:{s}^{2}\) is the sample variance of the paired differences and \(\:\frac{160}{483}\) is the test-to-training sample-size ratio, excluding the separate validation set. The test statistic was the mean paired difference divided by this standard error and was evaluated using a two-sided t test with 49 degrees of freedom. Holm correction was applied across the seven comparisons. Corrected 95% confidence intervals for mean AUC differences were not adjusted for multiplicity. Sensitivity, specificity, and accuracy were treated as descriptive secondary outcomes.
Results
A total of 803 TPUS recordings from 803 unique patients were included (Fig. 1). Median age was 58.0 years [IQR 47.9–68.6] overall, 58.5 years [IQR 49.5–68.7] in the SUI group, and 57.4 years [IQR 46.1–68.5] in the group without SUI (Table 1). The 95 videos selected for segmentation annotation had a median patient age of 53.3 years [IQR 41.6–62.2] and an SUI prevalence of 33.7%.
The automated segmentation check found nonempty bladder and urethra predictions in all 803 videos. No symphysis pubis voxels were predicted in five videos (0.6%). Visual review of these five videos confirmed that the symphysis pubis was outside the ultrasound field of view. Segmentation output was sufficient to construct a crop in every video, and none required fallback to the full frame. Minimum-size crop expansion was required in four videos (0.5%) in the entire cohort, all within the five videos without a visible symphysis pubis.
Mean test AUCs ranged from 0.84 to 0.90 (Table 2). V-JEPA2 exceeded USCL + TF in the basic configuration by a mean AUC difference of 0.052 (corrected 95% CI 0.019–0.084, Holm-adjusted p = 0.017). None of the other six paired AUC comparisons was statistically significant after Holm correction (Table 3).
Explainability analyses suggested that both models relied on anatomically and temporally relevant patterns in the TPUS examinations. For USCL + TF, Grad-CAM30 was applied on the frame encoder to highlight the most salient parts of the video frames for the classification decision of the model. In the reviewed examples, the resulting highlighted regions were typically close to the bladder neck across examination phases and extended toward the urethra during coughing as exemplified in Fig. 3. Similar patterns were seen in the crop and crop+append configurations, although some cases showed spatial drift or frame-to-frame flickering.
For V-JEPA2, spatial attention was more diffuse and less consistently localized to discrete anatomic landmarks. Temporal attention profiles were therefore reviewed qualitatively across multiple sequences. Figure 4 was selected as an illustrative example of the observed pattern, with an attention peak during Valsalva. As cohort-wide phase annotations were unavailable, the consistency of this pattern across the full cohort could not be quantified. Overall, these analyses indicate that both models captured plausible spatial or temporal patterns, while remaining qualitative and supportive rather than mechanistic.
Illustrative Grad-CAM visualizations for the USCL + TF model across TPUS examination phases: (A) rest, (B) pelvic floor contraction, (C) Valsalva maneuver, and (D) coughing. For each phase, the original ultrasound frame is shown next to the corresponding Grad-CAM overlay. Brighter regions indicate image areas with greater relevance for the model’s SUI classification decision. Saliency is primarily localized around the bladder neck and proximal urethral region, with more extensive activation during coughing.
Example temporal attention profile of the V-JEPA2 classification probe across a 64-frame TPUS sequence. Attention values were averaged per frame and normalized to the maximum value within the sequence. Shaded regions indicate the retrospectively annotated examination phases: rest, pelvic floor contraction, Valsalva maneuver, and coughing. In this example, attention peaks during the Valsalva phase.
Discussion
This study compared two deep-learning strategies for automated classification of stress urinary incontinence (SUI) from complete midsagittal transperineal ultrasound (TPUS) videos: a frame-wise ultrasound-pretrained approach with transformer-based temporal aggregation (USCL + TF) and a video-native foundation model (V-JEPA2). Both model families showed good discriminative performance, with mean AUCs ranging from 0.84 to 0.90. The highest mean AUCs were observed for V-JEPA2 in the crop and crop+append settings (both 0.90). Across matched input configurations, V-JEPA2 showed numerically higher mean AUC values than USCL + TF. The difference was statistically supported in the basic configuration, while the crop and crop+append differences were not significant after correction for multiple comparisons. The mean AUC of 0.89 for the basic V-JEPA2 model shows that full-video TPUS classification without manual frame or region-of-interest selection is feasible in this setting and does not depend on segmentation-guided input.
One plausible explanation for the numerically higher AUCs of V-JEPA2 is that large-scale video pretraining provides a stronger initialization for tasks that depend on temporal dynamics and motion-related information. In contrast, although USCL + TF benefits from ultrasound-specific pretraining at the frame level, temporal relationships are modeled only by a comparatively lightweight downstream transformer trained on a much smaller task-specific dataset. The observed differences may therefore reflect a combination of pretraining source, architecture, and temporal modeling. The anatomy-guided configurations showed numerical differences, but neither crop nor crop+append produced a statistically supported improvement over basic input within either architecture.
Although the anatomy-guided inputs did not produce statistically supported gains, the segmentation model performed broadly in line with prior work. Our dice scores of 0.85, 0.82, and 0.70 for the bladder, urethra, and symphysis pubis were comparable with corresponding values of 0.87, 0.61, and 0.55 reported by Szentimrey et al.17 using 248 midsagittal 3D TPUS volumes. This comparison should be interpreted with caution because the datasets and annotation protocols differed. For crop construction, masks were combined across the complete video, so an isolated poor mask was less likely to exclude relevant anatomy when other frames localized the structures adequately.
The explainability analyses provide some qualitative support that both model families used clinically plausible information. However, Grad-CAM and probe attention examine different model components and do not provide a direct comparison between the architectures. Cohort-wide phase timestamps were also unavailable, so the consistency of the observed Valsalva pattern could not be quantified. These findings therefore remain supportive rather than mechanistic.
Earlier comparable studies primarily relied on frame-based workflows with manual frame selection and, in some cases, manual ROI definition18,19,20,21. Chen et al. reported an AUC of 0.87 using manually selected Valsalva frames and manually annotated urethral ROIs21, and Zhang et al. reported an AUC of 0.92 using selected rest and Valsalva images20. By using complete video sequences after adequate urethral visualization had been confirmed, the present approach avoided these task-specific selection steps. This may facilitate future workflow integration, although its effects on processing time, staff time, cost, and clinical workload remain to be evaluated prospectively. Direct comparison across studies remains limited by differences in cohort composition, diagnostic definitions, reference standards, and preprocessing workflows.
Beyond reducing task-specific input selection, a further motivation for this work is the potential to support more standardized interpretation. Majida et al.14 reported interobserver ICCs ranging from 0.61 to 0.96 when two blinded observers analyzed the same stored 3D and 4D TPUS volumes, depending on the structure and maneuver. Although that study assessed pelvic-floor measurements rather than SUI diagnosis, it shows that variation can arise during analysis even when acquisition is held constant. The model developed here provides a basis for comparing unaided and AI-assisted interpretation and determining whether automated analysis improves agreement among expert readers.
Several limitations related to the study design and model development should be considered. First, this was a retrospective single-center study without external validation, and all examinations were acquired on ultrasound systems from a single manufacturer. The findings therefore do not yet establish robustness across institutions, operators, or devices. Second, although the cohort is larger than in several prior reports, it remains moderate for deep-learning development, particularly for training and comparing multiple model configurations. Third, segmentation for the crop and crop+append pipelines depended on annotations from a limited subset and primarily one annotator, which may have introduced annotation bias. Despite random selection, the annotated subset was younger and had a lower SUI prevalence than the full cohort. Independent expert annotations were unavailable, and Dice scores were not measured in the remaining 708 videos. The automated segmentation check and targeted visual review did not establish anatomical accuracy across the full cohort. Fourth, the reference diagnosis was derived from routine clinical assessment rather than a study-specific adjudication process, which reflects real-world practice but may also introduce heterogeneity.
Further limitations relate to acquisition, video preparation, and the choice of operating point. The acquisition protocol specified a standardized sequence, but the recordings were obtained during routine clinical examinations and depended on patient participation. Because phase labels were not recorded, we could not retrospectively confirm that every video contained all four maneuvers in the intended order or quantify how well each maneuver was performed. Both architectures required a fixed input length of 64 frames. Consequently, the 119 videos (14.8%) containing more than 64 frames were downsampled using linear temporal sampling. This sampling retained coverage across the recorded sequence but reduced temporal resolution, so a brief event could have been missed. Brief events could also have been missed in videos with fewer frames because the mean acquisition frame rate was only 3.5 frames/s. Learning to disregard padded frames may add to the difficulty of temporal modeling, although this was not investigated separately. The pipeline also required adequate urethral visualization in the automatically extracted midsagittal video. Of the 940 recordings assessed at this stage, 137 (14.6%) were excluded on this basis. This does not necessarily mean that the full 4D volume was clinically unusable, because the urethra may have been visible in an adjacent plane. Finally, Youden’s J provided a descriptive operating point for this retrospective comparison. Any clinical threshold would need to be selected and validated for the intended use, considering the consequences of false-negative and false-positive results. Prospective multicenter studies with external test cohorts and standardized acquisition will be necessary to determine clinical utility and generalizability.
In conclusion, deep-learning models were able to classify SUI status from complete midsagittal TPUS videos with adequate urethral visualization and without manual frame or ROI selection. V-JEPA2 showed numerically higher mean AUCs across matched input configurations, with a statistically supported difference in the basic configuration. Segmentation-guided input refinement did not show a statistically supported improvement over basic input. These results support further evaluation of automated TPUS video analysis as a potential clinical decision-support tool and provide a basis for future multicenter validation.
Data availability
The datasets used during the current study are available from the corresponding author on reasonable request, subject to institutional approval.
References
Patel, P. D., Amrute, K. V. & Badlani, G. H. Pelvic organ prolapse and stress urinary incontinence: A review of etiological factors. Indian J. Urol. 23, 135–141. https://doi.org/10.4103/0970-1591.32064 (2007).
Peinado Molina, R. A., Martínez, H., Martínez Vázquez, A. & Martínez Galiano, J. M. Influence of pelvic floor disorders on quality of life in women. Front. Public. Health. https://doi.org/10.3389/fpubh.2023.1180907 (2023).
Wu, J. M. et al. Prevalence and Trends of Symptomatic Pelvic Floor Disorders in U.S. Women. Obstet. Gynecol. 123, 141–148. https://doi.org/10.1097/aog.0000000000000057 (2014).
Reddy, J. & Paraiso, M. F. Primary stress urinary incontinence: what to do and why. Rev. Obstet. Gynecol. 3, 150–155 (2010).
Lugo, T., Leslie, S. W., Mikes, B. A. & Riggs, J. Stress Urinary Incontinence, < (2024). https://www.ncbi.nlm.nih.gov/books/NBK539769/
Macura, K., Thompson, R., Bluemke, D. & Genadry, R. Magnetic resonance imaging in assessment of stress urinary incontinence in women: Parameters differentiating urethral hypermobility and intrinsic sphincter deficiency. World J. Radiol. 7, 394–404. https://doi.org/10.4329/wjr.v7.i11.394 (2015).
Osman, N. I., Li Marzi, V., Cornu, J. N. & Drake, M. J. Evaluation and Classification of Stress Urinary Incontinence: Current Concepts and Future Directions. Eur. Urol. Focus. 2, 238–244. https://doi.org/10.1016/j.euf.2016.05.006 (2016).
Bogusiewicz, M. Ultrasound imaging in urogynecology – state of the art 2016. Menopausal Rev. 3, 123–132. https://doi.org/10.5114/pm.2016.63060 (2016).
Wen, L., Zhao, B., Chen, W., Qing, Z. & Liu, M. Real-time assessment of the behaviour of the bladder neck and proximal urethra during urine leaking in the cough stress test (CST) in supine and standing positions using transperineal ultrasound. Int. Urogynecol. J. 31, 2515–2519. https://doi.org/10.1007/s00192-020-04273-w (2020).
4D Transperineal Ultrasound for the Diagnosis and Classification of Stress Urinary. Incontinence in Postmenopausal Women. J. Coll. Physicians Surg. Pakistan. 33, 438–442. https://doi.org/10.29271/jcpsp.2023.04.438 (2023).
Santoro, G. A. et al. State of the art: an integrated approach to pelvic floor ultrasonography. Ultrasound Obstet. Gynecol. 37, 381–396. https://doi.org/10.1002/uog.8816 (2011).
Huang, Y. L. & Chen, H. Y. Computer-aided diagnosis of urodynamic stress incontinence with vector-based perineal ultrasound using neural networks. Ultrasound Obstet. Gynecol. 30, 1002–1006. https://doi.org/10.1002/uog.4102 (2007).
Wang, J. et al. Deep learning–assisted two-dimensional transperineal ultrasound for analyzing bladder neck motion in women with stress urinary incontinence. Am. J. Obstet. Gynecol. https://doi.org/10.1016/j.ajog.2024.07.021 (2025).
Majida, M. et al. Interobserver repeatability of three- and four‐dimensional transperineal ultrasound assessment of pelvic floor muscle anatomy and function. Ultrasound Obstet. Gynecol. 33, 567–573. https://doi.org/10.1002/uog.6351 (2009).
De Marzi, L., Baumfeld, Y., Wei, Q. & Shobeiri, S. A. Review of Emerging Technologies in Pelvic Floor Ultrasonography. Open. J. Obstet. Gynecol. 12, 375–385. https://doi.org/10.4236/ojog.2022.125034 (2022).
Santoro, G. A. Imaging the pelvic floor. Tech. Coloproctol. 21, 497–499. https://doi.org/10.1007/s10151-017-1668-y (2017).
Szentimrey, Z. et al. Automated segmentation and measurement of the female pelvic floor from the mid-sagittal plane of 3D ultrasound volumes. Med. Phys. 50, 6215–6227. https://doi.org/10.1002/mp.16389 (2023).
Szabo, T. et al. Innovations in Stress Urinary Incontinence: A Narrative Review. Medicina 61, 1272. https://doi.org/10.3390/medicina61071272 (2025).
Wang, Q., Wang, X., Jiang, X. & Lin, C. Machine learning in female urinary incontinence: A scoping review. Digit. HEALTH. 10, 20552076241281450. https://doi.org/10.1177/20552076241281450 (2024).
Zhang, M. et al. Artificial intelligence models derived from 2D transperineal ultrasound images in the clinical diagnosis of stress urinary incontinence. Int. Urogynecol. J. 33, 1179–1185. https://doi.org/10.1007/s00192-021-04859-y (2022).
Chen, K. et al. An optimized deep learning model based on transperineal ultrasound images for precision diagnosis of female stress urinary incontinence. Front. Med. https://doi.org/10.3389/fmed.2025.1564446 (2025).
Russakovsky, O. et al. ImageNet large scale visual recognition challenge. Int. J. Comput. Vis. (IJCV). 115, 211–252. https://doi.org/10.1007/s11263-015-0816-y (2015).
Isensee, F., Jaeger, P. F., Kohl, S. A. A., Petersen, J. & Maier-Hein, K. H. nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation. Nat. Methods. 18, 203–211. https://doi.org/10.1038/s41592-020-01008-z (2021).
Arnab, A. et al. ViViT: A Video Vision Transformer. 2021 IEEE/CVF Int. Conf. Comput. Vis. (ICCV). 6816–6826. https://doi.org/10.1109/iccv48922.2021.00676 (2021).
Neimark, D., Bar, O., Zohar, M. & Asselmann, D. Video Transformer Network. 2021 IEEE/CVF Int. Conf. Comput. Vis. Workshops (ICCVW). 3156–3165. https://doi.org/10.1109/iccvw54120.2021.00355 (2021).
Chen, Y. et al. Pretraining deep ultrasound image diagnosis model through video contrastive representation learning. Lect Notes Comput. Sci. https://doi.org/10.1007/978-3-030-87237-3_60 (2021).
He, K., Zhang, X., Ren, S. & Sun, J. Deep Residual Learning for Image Recognition. IEEE Conf. Comput. Vis. Pattern Recognit. https://doi.org/10.1109/CVPR.2016.90 (2016).
Assran, M. et al. V-JEPA 2: Self-supervised video models enable understanding, prediction and planning. Preprint at https://arXiv.org/abs/2506.09985.
Nadeau, C. & Bengio, Y. Inference for the Generalization Error. Mach. Learn. 52, 239–281. https://doi.org/10.1023/a:1024068626366 (2003).
Selvaraju, R. R. et al. Grad-CAM: Visual explanations from deep networks via gradient-based localization. 2017 IEEE Int. Conf. Comput. Vis. (ICCV). https://doi.org/10.1109/iccv.2017.74 (2017).
Acknowledgements
The authors acknowledge Martin Mienkina and Dimitrios Kastanis at GE HealthCare for providing access to software enabling conversion of proprietary imaging data into NRRD format and for technical support in its use.
Funding
Open Access funding enabled and organized by Projekt DEAL. DT is supported by the German Ministry of Research, Technology and Space (TRANSFORM LIVER − 031L0312C, DECIPHER-M − 01KD2420B), Deutsche Forschungsgemeinschaft (DFG) (515639690), and the European Union (Horizon Europe, ODELIA - GA 101057091, ERC Starting Grant SAGMA – GA 101222556).
Author information
Authors and Affiliations
Lab for Artificial Intelligence in Medicine, Department of Diagnostic and Interventional Radiology, University Hospital Aachen, Aachen, Germany
Roman Vuskov, Nikol Ignatova & Daniel Truhn
Department of Diagnostic and Interventional Radiology, University Hospital Aachen, Aachen, Germany
Roman Vuskov & Daniel Truhn
Department of Gynecology and Obstetrics, University Hospital Aachen, Aachen, Germany
Laila Najjari
Authors
- Roman Vuskov
- Nikol Ignatova
- Daniel Truhn
- Laila Najjari
Contributions
RV conducted the experiments and analysis, wrote the initial manuscript text, and prepared all figures. NI performed data preparation and segmentations. LN conducted medical data quality control and provided clinical oversight. DT provided scientific supervision. All authors reviewed and approved the final manuscript.
Corresponding author
Correspondence to Roman Vuskov.
Ethics declarations
Competing interests
DT received honoraria for lectures by Bayer, GE, Roche, Astra Zeneca, and Philips and holds shares in StratifAI GmbH, Germany and in Synagen GmbH, Germany. RV, NI and LN declare no potential conflict of interest.
Additional information
Publisher’s note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Rights and permissions
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.
About this article
Cite this article
Vuskov, R., Ignatova, N., Truhn, D. et al. Deep learning–based detection of stress urinary incontinence in transperineal ultrasound video examinations. Sci Rep 16, 37982 (2026). https://doi.org/10.1038/s41598-026-77927-2
Received: 12 June 2026
Accepted: 22 October 2026
Published: 14 November 2026
Version of record: 14 November 2026
DOI: https://doi.org/10.1038/s41598-026-77927-2
来源:Nature 健康护理主题 · nature.com