A lightweight ResNet50V2-ECA model for renal cell carcinoma grading: efficiency, calibration, and state-of-the-art...

Accurate International Society of Urological Pathology (ISUP)-grade classification of renal cell carcinoma (RCC) is challenging due to subtle histopathological variations and the limitations of manual review. Existing deep learning models often rely on complex attention mechanisms that increase computational cost and hinder deployment. This study introduces a lightweight ResNet50V2-ECA framework that adds a single Efficient Channel Attention block after the final convolutional layer, enhancing channel-wise discrimination with minimal overhead (0.25 M parameters, 3.1 GFLOPs). Placing a single ECA block before GAP allows global channel recalibration after spatial feature extraction, avoiding the computational redundancy of multi-stage attention while preserving spatial resolution. Using the KMC kidney histopathology dataset (722 WSIs, 3,442 training patches across five grades), the model achieves 96.90% accuracy, 91.39% F1-score, and near-perfect AUC scores (Grade-0: 1.000; Grade-4: 0.997) across standard metrics (accuracy, precision, recall, F1, specificity, BAC, AUC). A comprehensive ablation study across five attention mechanisms (BAM, CBAM, SE, GC, ECA) and three backbones confirms that ResNet50V2-ECA yields the highest overall performance. Comparative analysis with leading RCC frameworks (RoCNN, EFF-Net, RCCGNet, RenalNet, MobileDANet) further demonstrates superior accuracy and efficiency. Model reliability is validated through a calibration diagram using Monte Carlo Dropout, showing strong alignment between predicted confidence and actual accuracy, while the risk-coverage curve reveals accuracy exceeding 99% when low-confidence cases are deferred. Together, these findings establish a high-precision, well-calibrated RCC grading system with strong potential for clinical deployment. While results indicate strong potential for clinical deployment, external multi-center validation is required to confirm generalizability.