Journal of pediatric urology · 2026
CurvAssist: An AI assisted pipeline for penile shaft segmentation/ curvature measurement in children with hypospadias.
СтатьяНазвание и аннотация: Перевод в очереди
Обработка статьи · этапы, ошибки и управление
Загрузка состояния…
Ошибки и ограничения
История · было → стало
Состояние обновляется каждые 5 секунд, пока блок открыт. Процент показывается только там, где известен объём работы. Пауза останавливает продолжение этапа; текущий запрос к источнику может завершиться.
Babu Ramesh, A Sathya, Sivakumar Deena, Vijayaraghavan Geetha, Velmurugan Maniselvam, Sampathkumar Kabilash, Shakthi Dheepak
Аннотация
INTRODUCTION: Accurate assessment of ventral curvature (VC) is critical in determining the type of VC correction. Unaided visual inspection (UVI) remains common but is prone to significant interobserver variability and error. Although smartphone app-based measurement (SAM) offers greater objectivity, it still relies on manual reference point selection. This study describes the development and evaluation of CurvAssist, an AI-assisted pipeline integrating deep learning-based segmentation for accurate, automated penile curvature measurement.
METHODS: In step-1, a synthetic dataset of 250 images was generated from 3D phantom penile models with known ground-truth curvatures (0°-90°) under randomized lighting, angle, and background conditions. Penile shaft segmentation was achieved using a U-Net convolutional neural network. Curvature quantification employed a nonlinear least-squares circle-fitting algorithm using the Levenberg-Marquardt optimization to fit arc contours and calculate curvature angles geometrically. In step-2, the circle fitting algorithm's outputs were compared against SAM on clinical photographs (n = 7). Model performance was evaluated using Dice Similarity Coefficient (DSC), Intersection over Union (IoU), Precision, Recall, Mean Absolute Error (MAE), and Intraclass Correlation Coefficient (ICC).
RESULTS: U-Net that analysed phantom 3D images (n = 250) achieved high segmentation accuracy with mean DSC 0.925 ± 0.021, IoU 0.910 ± 0.035, Precision 0.958, and Recall 0.947, providing a robust foundation for geometric quantification. The circle-fit algorithm estimated curvature with a MAE of 1.35°, representing a clinically negligible deviation from ground-truth. Comparison of AI-derived versus SAM-measured angles of clinical photographs (n = 7) demonstrated excellent agreement (ICC = 0.981).
DISCUSSION: CurvAssist achieved a notable improvement and compares favourably with other emerging computational methods. While the AI-based applications have evolved to match the level of SAM measurements, they are likely to be more consistent when the model becomes robust avoiding human errors/relying on manually selected reference points on SAM. CurvAssist should still be considered as a proof-of concept or technical feasibility study requiring further large-scale validation. The time required for segmentation and computation was not evaluated in this study.
CONCLUSION: CurvAssist can provide surgeons with objective, reproducible data to guide intraoperative decision-making, standardize surgical planning, and enhance multi-center research by minimizing interobserver variability. Future research should focus on prospective validation using a large dataset of intraoperative images, expanding the system to perform true 3D measurements from multiple views, and in developing a user-friendly real-time interface for seamless integration into the clinical workflow while addressing regulatory concerns on data handling and privacy.
METHODS: In step-1, a synthetic dataset of 250 images was generated from 3D phantom penile models with known ground-truth curvatures (0°-90°) under randomized lighting, angle, and background conditions. Penile shaft segmentation was achieved using a U-Net convolutional neural network. Curvature quantification employed a nonlinear least-squares circle-fitting algorithm using the Levenberg-Marquardt optimization to fit arc contours and calculate curvature angles geometrically. In step-2, the circle fitting algorithm's outputs were compared against SAM on clinical photographs (n = 7). Model performance was evaluated using Dice Similarity Coefficient (DSC), Intersection over Union (IoU), Precision, Recall, Mean Absolute Error (MAE), and Intraclass Correlation Coefficient (ICC).
RESULTS: U-Net that analysed phantom 3D images (n = 250) achieved high segmentation accuracy with mean DSC 0.925 ± 0.021, IoU 0.910 ± 0.035, Precision 0.958, and Recall 0.947, providing a robust foundation for geometric quantification. The circle-fit algorithm estimated curvature with a MAE of 1.35°, representing a clinically negligible deviation from ground-truth. Comparison of AI-derived versus SAM-measured angles of clinical photographs (n = 7) demonstrated excellent agreement (ICC = 0.981).
DISCUSSION: CurvAssist achieved a notable improvement and compares favourably with other emerging computational methods. While the AI-based applications have evolved to match the level of SAM measurements, they are likely to be more consistent when the model becomes robust avoiding human errors/relying on manually selected reference points on SAM. CurvAssist should still be considered as a proof-of concept or technical feasibility study requiring further large-scale validation. The time required for segmentation and computation was not evaluated in this study.
CONCLUSION: CurvAssist can provide surgeons with objective, reproducible data to guide intraoperative decision-making, standardize surgical planning, and enhance multi-center research by minimizing interobserver variability. Future research should focus on prospective validation using a large dataset of intraoperative images, expanding the system to perform true 3D measurements from multiple views, and in developing a user-friendly real-time interface for seamless integration into the clinical workflow while addressing regulatory concerns on data handling and privacy.