Bridging explainable machine vision in CAD Systems for lung cancer detection

Computer-aided diagnosis (CAD) systems have grown increasingly popular with aiding physicians in diagnosing lung cancer using medical images in recent years. However, the reasoning behind the state-of-the-art black-box learning and prediction models has become obscured and this resultant lack of transparency has presented a problem whereby physicians are unable to trust the results of these systems. This motivated us to improve the conventional CAD with a more robust and interpretable algorithms to produce a system that achieves high accuracy and explainable diagnoses of lung cancer. The proposed approach uses a novel image processing pipeline to segment nodules from lung CT scan images, and then classifies the nodule using both 2D and 3D Alexnet models that have been trained on lung nodule data from the LIDC-IDRI dataset. The explainability aspect is approached from two angles: 1) LIME that produces a visual explanation of the diagnosis, and 2) a rule-based system that produces a text-based explanation of the diagnosis. Overall, the proposed algorithm has achieved better performance and advance the practicality of CAD systems.