Explainable active reinforcement deep learning improves lung cancer detection from CT images

Abstract

Lung cancer remains a global health challenge that requires early and accurate diagnosis through medical imaging analysis. This study introduces ARXAF-Net framework which integrates Active Reinforcement deep leaning with strategic feature engineering, selection, advanced classification techniques with Explainable AI. Firstly, The ARXAF-Net framework overcomes the challenges of labeled dataset limitation by leveraging active reinforcement learning where the model achieves a remarkable 99.0% training accuracy using reinforcement learning. After that, Traditional feature extraction techniques, including GLCM and LBP, are combined with CNN with attention fusion model features, forming comprehensive vectors. Advanced techniques pinpoint essential characteristics for classification. The experiments conducted on a dataset comprising 30,020 CT images categorized into two classes—15,010 non-cancer and 15,010 cancer—demonstrate that CNN models employing attention fusion with traditional feature extraction achieve remarkable consistency, reaching a testing accuracy of 99%. Basic CNN models with traditional feature extraction, following normalization, display commendable performance, nearing an accuracy of 95%. Additionally, integrating explainable AI (XAI) into the performance frameworks significantly enhances the outcomes by incorporating feedback from radiologists. This research offers valuable insights into the optimal combinations of preprocessing, feature engineering/selection, and classification algorithms aimed at maximizing lung cancer detection efficacy. It also recognizes the trade-offs between accuracy and efficiency when merging deep and traditional features, highlighting the importance of careful feature selection. Moreover, addressing the challenges in this integration and investigating hyper-parameter tuning for machine learning models may present avenues for future improvements.