Scholarlink

<p>Kidney stones are a very common urological condition resulting in significant healthcare costs. It requires an accurate and timely diagnosis for the treatment to be effective. While CNNs exhibit high diagnostic accuracy, their clinical deployment is often hindered by high computational overhead and its "black box" nature resulting in lack of interpretability. This study benchmarks five pre-trained CNN models (ResNet50, InceptionV3, MobileNetV2, DenseNet121, EfficientNet-B0) for classifying kidney stone cases versus normal cases. Evaluating across a clinically balanced distribution of 1,787 normal and 1,577 stone-positive scans, InceptionV3 achieved the highest performance across all metrics: accuracy 98.66%, precision 0.9867, recall 0.9865, F1-score 0.9866, ROC-AUC 0.9969 and Expected Calibration Error (ECE) of 0.0091. This was followed by EfficientNet-B0 and MobileNetV2 achieving accuracies of 98.22% and 98.07% respectively, while ResNet50 and DenseNet121 achieved a similar accuracy score of 97.92%. Our findings demonstrate that lightweight architectures like EfficientNet-B0 and MobileNetV2 achieve statistically comparable performance to heavier models while being computationally efficient. From the study, it can be hypothesized that this parity may be due to their efficient depthwise processing of localized high density patterns in CT slices, which seems to relate well with kidney stone morphology compared to the broader feature mixing that are used by heavier architectures. Furthermore, enhanced Grad-CAM visualizations suggest that these networks can be repeatedly used to localize pertinent anatomical areas, which implies that they can be utilized in the clinical environment with limited resources. Future work should extend this framework to multi-class renal pathology classification, multi-centre validation, and prospective radiologist-in-the-loop evaluation.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Explainable deep learning models for kidney stone classification using axial CT scans

  • 期刊论文

  • 发表日期：

    2026年8月4日

查看全文
  • Muhammad Zeeshan Burki,
  • Rashid Naseem,
  • Abdul Wajahat,
  • Muhammad Rizwan,
  • Nazeer Muhammad

摘要

Kidney stones are a very common urological condition resulting in significant healthcare costs. It requires an accurate and timely diagnosis for the treatment to be effective. While CNNs exhibit high diagnostic accuracy, their clinical deployment is often hindered by high computational overhead and its "black box" nature resulting in lack of interpretability. This study benchmarks five pre-trained CNN models (ResNet50, InceptionV3, MobileNetV2, DenseNet121, EfficientNet-B0) for classifying kidney stone cases versus normal cases. Evaluating across a clinically balanced distribution of 1,787 normal and 1,577 stone-positive scans, InceptionV3 achieved the highest performance across all metrics: accuracy 98.66%, precision 0.9867, recall 0.9865, F1-score 0.9866, ROC-AUC 0.9969 and Expected Calibration Error (ECE) of 0.0091. This was followed by EfficientNet-B0 and MobileNetV2 achieving accuracies of 98.22% and 98.07% respectively, while ResNet50 and DenseNet121 achieved a similar accuracy score of 97.92%. Our findings demonstrate that lightweight architectures like EfficientNet-B0 and MobileNetV2 achieve statistically comparable performance to heavier models while being computationally efficient. From the study, it can be hypothesized that this parity may be due to their efficient depthwise processing of localized high density patterns in CT slices, which seems to relate well with kidney stone morphology compared to the broader feature mixing that are used by heavier architectures. Furthermore, enhanced Grad-CAM visualizations suggest that these networks can be repeatedly used to localize pertinent anatomical areas, which implies that they can be utilized in the clinical environment with limited resources. Future work should extend this framework to multi-class renal pathology classification, multi-centre validation, and prospective radiologist-in-the-loop evaluation.