Explaining Fuhrman Grading using Deep Convolutional-Driven Features of Renal Cell Carcinoma

Renal cell carcinoma (RCC) is the most prevalent kind of kidney cancer, making up around 80% of cases worldwide and showing signs of increasing incidence. Whole-slide imaging (WSI) has revolutionized computer-aided diagnosis through its usage in the recent deep learning (DL)-based studies, which assists pathologists in histologically grading RCC tumours and selecting the most effective treatment plans for patients. Despite DL’s effectiveness, its black-box nature limits transparency. To address this, a novel explainable framework is proposed for RCC histological grading, utilizing features extracted by a custom-built convolutional neural network (CNN). Clear cell RCC (ccRCC) images from The Cancer Genome Atlas (TCGA) were used, labelled with Fuhrman grades (FG) 1 to 4. Further, to analyse the WSI images, we extracted histological patches of 256 x 256 pixels at 40x magnification. These patches were classified into two-tiered (Grades 1/2-3/4) and four-tiered (Grades 1-4) FG by extracting deep features from histopathology patches through CNNs, subsequently fed into multiple machine learning classifiers. The experimental results demonstrated the robustness of the proposed framework through 5-fold cross-validation, achieving accuracies of 0.9731 ± 0.0095 and 0.9794 ± 0.0065 for two-tiered FG and four-tiered FG, respectively, surpassing the performance of existing models. Finally, the paper utilized explainable AI tool, Grad-CAM, to identify regions influencing specific grades, offering valuable clinical insights.