Interpretable rough neural network for lung nodule diagnosis

Abstract

Computer-aided diagnosis (CAD) systems based on deep learning have shown significant potential in lung nodule diagnosis, providing substantial assistance to medical professionals. However, the inherent lack of interpretability in deep learning models and the uncertainty of annotations limit their widespread application. We propose that uncertain annotations actually imply additional valuable information that can enhance both model performance and interpretability. To address these challenges, we have developed a novel soft computing methodology integrating rough sets with deep neural networks. Firstly, this methodology employs rough sets to process uncertain region of interest (ROI) annotations into upper and lower approximations. Secondly, a novel rough neuron is designed to predict these approximations. Thirdly, the newly proposed region-constraint strategy embeds interpretable radiological domain knowledge into the neural network. Finally, this methodology proposes interpretation curves and regional consistency metrics to quantitatively evaluate the model’s interpretability. We conducted extensive comparison experiments on LIDC-IDRI and LNDb public benchmarks. Detailed experimental results demonstrate that by maximally retaining uncertain samples, the proposed method achieves classification accuracies of 84.6% and 89.74%, and mean absolute errors of 0.4988 and 0.5208 in attribute prediction, representing improvements of 3.4% and 2.5%, respectively, over the backbone networks.