An ISHAP-based interpretation-model-guided classification method for malignant pulmonary nodule

Abstract

The classification of benign and malignant pulmonary nodules can provide an important aid to the diagnosis of lung cancer. However, high-performance classification models are still less accepted approaches in clinic applications due to their structural complexity and low interpretability. In this study, an ISHAP (Improved SHapley Additive exPlanations)-based interpretation-model-guided classification method is proposed for the classification of benign and malignant pulmonary nodules. First, semantic and radiomics features are extracted by using medical priori knowledge and image understanding in the lung images dataset. Then, optimal features, classifiers and their parameters are guided to select and set adaptively based on the proposed ISHAP explanation and recursive feature elimination algorithm. The proposed ISHAP-based interpretation-model-guided classification method on the Lung Image Database Consortium (LIDC) dataset achieves a sensitivity of 0.862, a specificity of 0.885, an accuracy of 0.873 and an area under the receiver-operating-characteristic curve of 0.941 on testing dataset. Meanwhile, our analysis with ISHAP-based interpretation model can reveal the essential factors of semantic and radiomics features in the classification of malignant pulmonary nodules. Experimental results demonstrate that the proposed method achieves satisfactory classification performance and makes the predictions more transparent and trustworthy by interpreting the model.