An Explainable Classification Model of Renal Cancer Subtype Using Deep Learning

Cancer diagnosis often presents significant challenges for doctors, as it encompasses multiple types of disease. Ultrasound renal images suffer from low resolution, low contrast and high speckle noise. There is a need for an effective model to accurately classify US renal cancer images and locate lesions within them. In this paper, we propose a new explainable deep learning model for cancer image classification, which consists of a Deep Learning model and a Gradient-Weighted Class Activation Mapping (Grad-CAM). This model can distinguish between hamartoma and clear cell renal cell carcinoma, as well as accurately locate the lesion area. Experimental results demonstrate that our method significantly improves model performance and the area under the curve in comparison to other general models such as Vgg16 and Resnet34 using Grad-CAM. By combining Grad-CAM, our model can more accurately locate lesion areas and is robust to noise interference. Superior performance shown the potential of the proposed method as a sensitive classification tool that may help develop AI-based computer-aided diagnosis.