人工智慧在水腎症識別中的應用：靜脈腎盂造影影像的評估

黃東亮¹、林恩²、郭永明²、盧則宏³、劉展榮³

國立成功大學醫學院附設醫院 教學中心¹，泌尿部³；國立虎尾科技大學 電子工程系²

Artificial Intelligence for Hydronephrosis Recognition: An Evaluation on Intravenous Pyelography Images

Tung-Liang Hwang¹, En Lin², Yung-Ming Kuo², Ze‐Hong Lu³, Chan-Jung Liu³

Education Center¹, Department of Urology³, National Cheng Kung University Hospital, College of Medicine, National Cheng Kung University, Tainan, Taiwan; Department of Electronic Engineering², National Formosa University, Yunlin, Taiwan

Purpose: Intravenous pyelography (IVP) remains the gold standard for evaluating urinary tract anatomy and function, yet hydronephrosis diagnosis via IVP is highly subjective. This study aimed to develop and validate a deep learning model to enhance reliability in detecting hydronephrosis from IVP images.

Materials and Methods: In this retrospective study (IRB: A-ER-114-207), IVP images were collected between January 2017 and December 2024 from our institution as an internal dataset for training. Two clinicians’ experts were invited to identify the truth of the internal datasets, which were endourologist (C.-J. L. and T.-L. H. with 10 years of experience reading IVP radiographs, respectively). Readers were blinded to clinical information and other imaging results. The current training sample size was n=103. Samples were divided into training (N = 57) and validation sets (N = 46). An ensemble learning-based model, using Res-Net 18, VGG16, VGG19, and DenseNet121. The grading performance was assessed using Youden, sensitivity, specificity, and F1 score. The detailed methods were the follows: Images were categorized into left and right sides. Right-sided images were mirrored to left-side orientation and combined with original left-sided images to form a full left-side dataset, the same as left-side. Finally, the full left- and right-side datasets were combined to produce three final data sets (left, right, combined left and right). Leave-one-out cross-validation (LOOCV) was performed on each dataset.

Results: The outcomes based on the area under curve (AUC), sensitivity, specificity, Youden index, and F1-score. For the left-side datasets, the best model was VGG19, which achieved AUC of 0·814 [95% confidence interval (CI) 0·733–0·896], sensitivity of 0.930 [95% CI 0.927–0.933], specificity of 0.880 [95% CI 0.872–0.888], Youden of 0.810 [95% CI 0.802–0.818], and F1 of 0.920 [95% CI 0.917–0.923]. The worst model was DenseNet121. For the right-side datasets, the best model was also VGG19. VGG19 can achieve AUC of 0·815 [95% CI 0·734–0·895], sensitivity of 0.880 [95% CI 0.877–0.883], specificity of 0.900 [95% CI 0.887–0.913], Youden of 0.780 [95% CI 0.757–0.803], and F1 of 0.900 [95% CI 0.897–0.903]. However, for the combination dataset, DenseNet121 performed best: AUC 0·709 [95% CI 0·640–0·779], sensitivity of 0.960 [95% CI 0.958–0.962], specificity of 0.950 [95% CI 0.947–0.953], Youden of 0.900 [95% CI 0.895–0.905], and F1 of 0.960 [95% CI 0.958–0.962].

Conclusions: The Deep-IVP model demonstrates high accuracy and objectivity for hydronephrosis detection on IVP, providing a generalizable tool to support radiographic evaluation.