<p>The purpose of this study was to investigate the effectiveness of deep learning (DL) methods in recognizing uric acid types from images of urinary stones. We curated a single-center cohort of 208 anonymized CT images from patients with urinary calculi. Images were rigorously screened (single calculus, diameter greater than or equal to 3 mm, high quality without artifacts), uniformly preprocessed to 512 × 512, and selected at the patient level (one image per patient) to prevent data leakage. Stones were labeled by Hounsfield Units from clinical reports into three classes—uric acid (0-600 HU), mixed (600–1000 HU), and non-uric acid (greater than 1000 HU)—yielding 50/72/86 images, respectively. We trained and validated these 208 urinary stone images using 7 deep learning (DL) networks and compared classification efficiency. Among seven deep learning models evaluated on the same curated dataset (<i>n</i> = 208), ResNet18 yielded the best performance (accuracy = 0.9808, precision = 0.9828, recall = 0.9761, F1 = 0.9790) and was selected as the final model for urinary stone composition prediction. At the argmax point, class-wise recalls were 94.00% (UA), 98.84% (NUA), and 100% (MIX); corresponding AUCs were 0.984, 0.987, and 0.977 (macro AUC = 0.985), and APs were 0.974, 0.983, and 0.952 (macro AP = 0.970). Sensitivity/specificity reached 0.94/1.00 (UA), 0.9884/0.9918 (NUA), and 1.00/0.9779 (MIX), indicating no missed MIX cases and no false positives for UA. Decision curve analysis (DCA) and clinical impact curves (CIC) showed positive net benefit across clinically plausible thresholds (notably around <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(P_t=0.4\)</EquationSource> </InlineEquation> ), suggesting efficient capture of true cases with a manageable screening burden. This study suggests that a ResNet18-based deep learning model may support highly accurate, preoperative, noninvasive identification of uric acid stones and has potential clinical utility. The approach has the potential to streamline the diagnostic-therapeutic pathway for urolithiasis and to reduce unnecessary invasive procedures. Future multicenter, prospective investigations will evaluate its real-world generalizability and cost-effectiveness, and explore multimodal data fusion to further improve recognition of complex stones. Our code and datasets are available at <a href="https://github.com/wbin2002/UrinaryStone.git">https://github.com/wbin2002/UrinaryStone.git</a>.</p>

Recognizing uric acid type of urinary stones by deep learning

期刊论文
发表日期：

2025年12月26日

查看全文

Bin Wang,
Yingying Zhou,
Jiaxin Cai

摘要

The purpose of this study was to investigate the effectiveness of deep learning (DL) methods in recognizing uric acid types from images of urinary stones. We curated a single-center cohort of 208 anonymized CT images from patients with urinary calculi. Images were rigorously screened (single calculus, diameter greater than or equal to 3 mm, high quality without artifacts), uniformly preprocessed to 512 × 512, and selected at the patient level (one image per patient) to prevent data leakage. Stones were labeled by Hounsfield Units from clinical reports into three classes—uric acid (0-600 HU), mixed (600–1000 HU), and non-uric acid (greater than 1000 HU)—yielding 50/72/86 images, respectively. We trained and validated these 208 urinary stone images using 7 deep learning (DL) networks and compared classification efficiency. Among seven deep learning models evaluated on the same curated dataset (n = 208), ResNet18 yielded the best performance (accuracy = 0.9808, precision = 0.9828, recall = 0.9761, F1 = 0.9790) and was selected as the final model for urinary stone composition prediction. At the argmax point, class-wise recalls were 94.00% (UA), 98.84% (NUA), and 100% (MIX); corresponding AUCs were 0.984, 0.987, and 0.977 (macro AUC = 0.985), and APs were 0.974, 0.983, and 0.952 (macro AP = 0.970). Sensitivity/specificity reached 0.94/1.00 (UA), 0.9884/0.9918 (NUA), and 1.00/0.9779 (MIX), indicating no missed MIX cases and no false positives for UA. Decision curve analysis (DCA) and clinical impact curves (CIC) showed positive net benefit across clinically plausible thresholds (notably around \(P_t=0.4\) ), suggesting efficient capture of true cases with a manageable screening burden. This study suggests that a ResNet18-based deep learning model may support highly accurate, preoperative, noninvasive identification of uric acid stones and has potential clinical utility. The approach has the potential to streamline the diagnostic-therapeutic pathway for urolithiasis and to reduce unnecessary invasive procedures. Future multicenter, prospective investigations will evaluate its real-world generalizability and cost-effectiveness, and explore multimodal data fusion to further improve recognition of complex stones. Our code and datasets are available at https://github.com/wbin2002/UrinaryStone.git.