Explainable Prediction Of Renal Cell Carcinoma From Contrast-Enhanced CT Images Using Deep Convolutional Transfer Learning And The Shapley Additive Explanations Approach

• Main Authors: Han, Fuchang, Liao, Shenghui, Yuan, Siming, Wu, Renzhong, Zhao, Yuqian, Xie, Yu
• Format: Conference Proceeding
• Language: English
• Subjects: Additives; Analytical models; Computed tomography; Extra trees; Hospitals; Malignant tumors; Predictive models; Renal cell carcinoma; ResNet-101; SHapley Additive exPlanations; Transfer learning
• ISSN: 2381-8549
• DOI: 10.1109/ICIP42928.2021.9506144

The prediction of renal cell carcinoma (RCC) is an important cancer screening step. Existing state-of-the-art methods focus on developing machine/deep learning networks with one or more optimization strategies for higher identification accuracies. Such developments ignore the interpretability and clinical utility of models, which are still quite opaque to clinicians. This paper introduces deep convolutional transfer learning and SHapley Additive exPlanations (SHAP) to the classification model and proposes an explainable RCC prediction model. The model evaluates the risks and benefits using decision curve analysis (DCA). Specifically, multiscale feature extraction and compensation are proposed to enrich the representations. By combining the high-importance features in a parallel manner, the models' performances are gradually enhanced. Our model achieves an accuracy of 73.87% and an area under the curve (AUC) of 0.8030 on the Hunan Cancer Hospital dataset. To demonstrate the generalizability, our model yields an accuracy of 99.81% on the public COIL-100 dataset.