A Hybrid Radiomics–Deep Learning Feature-Fusion Framework with Explainable Artificial Intelligence for Accurate Liver Cancer Prediction from Computed Tomography Images

Abstract

Hepatocellular Carcinoma (HCC) is one of the common cancer-related mortality causes across the world. The black-box nature (semantic opacity) of current deep learning models used in computer-aided detection hinders clinical management of HCC, particularly the lack of early and interpretable diagnosis. In this work, a hybrid fusion framework at the feature level is developed to enhance the accuracy and interpretability of predicting liver cancer using computed tomography (CT) images. CT volumes dataset from the multi-center Liver Tumor Segmentation (LiTS) benchmark were pre-processed using Hounsfield-Unit windowing, isotropic resampling and automated, spatial-pyramid-pooling-based segmentation. A 3-D deep feature was extracted using a 3-dimensional ResNet-10 backbone and a 562-dimensional hybrid vector was created by concatenation PyRadiomics descriptors selected by Lasso, balanced with Synthetic Minority Over-sampling Technique (SMOTE), an eXtreme Gradient Boosting (XGBoost) model optimized by GridSearchCV was then used for classification. To obtain both spatial and feature-level explainability, Gradient-weighted Class Activation Mapping (Grad-CAM) and SHapley Additive exPlanations (SHAP) are applied. The AUC-ROC of the optimized hybrid classifier was 0.921 and an overall accuracy of 85% and sensitivity of 0.95 for low tumor-burden cases were obtained. Predictions were shown to be based on biologically relevant regions of the tumor, and not on anatomical noise by using Grad-CAM heatmaps. The results suggest that a radiomics–deep learning fusion model with explainability and transparency has high diagnostic accuracy and clinical interpretability, and can be used as a reliable model for decision support in the field of hepatic oncology.