Shapley Patch Valuation Method for Histopathological Image Classification

Abstract

This study introduces a patch valuation method based on Shapley values for histopathological image classification, addressing the computational challenges of processing large whole-slide images (WSIs). The proposed approach leverages Shapley values to identify the most informative patches, optimizing dataset size and computational efficiency. We evaluate the performance of three tree-based ensemble models - XGBoost, LightGBM, and CatBoost - on a subset of the EBHI dataset containing histopathological images of colorectal cancer captured at 200× magnification. Each image is divided into 128×128 patches, and Shapley values are computed to rank patch importance. The top patches are selected to train the models, and their performance is evaluated using accuracy, precision, recall, and F1-score. The results demonstrate that smaller patch sizes (e.g., 20K patches) achieve comparable performance (0.8049 vs 0.8374 and 80 vs 969s) to the full dataset (270K patches), with XGBoost and LightGBM showing balanced performance across metrics. CatBoost achieves the highest accuracy but requires significantly more training time. LightGBM proves to be the fastest model, making it ideal for scenarios prioritizing computational efficiency. The study highlights the effectiveness of Shapley value-based patch selection in reducing computational complexity while maintaining high classification accuracy. This approach offers significant implications for optimizing digital pathology workflows and improving the efficiency of histopathological image classification systems.