FVCM-Net: Interpretable privacy-preserved attention driven lung cancer detection from CT scan images with explainable HiRes-CAM attribution map and ensemble learning

Abstract

Lung cancer is a predominant cause of cancer-related deaths globally, and early detection is essential for improving patient prognosis. Deep learning models with attention mechanisms have shown promising accuracy in detecting lung cancer from medical imaging data. However, privacy concerns and data scarcity present significant challenges in developing robust and generalizable models. This paper proposes a novel approach for lung cancer detection, ‘FVCM-Net’, integrating federated learning with attention mechanisms and ensemble learning to address these challenges. Federated learning is employed to train the model across multiple decentralized institutions, allowing for collaborative model development without sharing sensitive patient data and minimizing the risk of such sensitive data being misused. Furthermore, this approach enables the development of more accurate and generalized models by leveraging diverse datasets from multiple sources. We employed ensemble learning to produce more accurate predictions than a single model. For interpretability of the lung cancer identification model, we employ XAI (Explainable Artificial Intelligence) techniques such as SHAP (SHapley Additive exPlanations) and HiResCAM (High-Resolution Class Activation Mapping). These techniques help us understand how the model makes its decisions and predictions. This study utilizes a diverse collection of lung CT scan images from four datasets, including LIDC-IDRI, IQ-OTH/NCCD, a public Kaggle dataset, and additional online sources. Experimental results revealed that the proposed method achieved higher performance in lung cancer detection with 98.26% average accuracy and 97.37% average F-1 score. The high performance of FVCM-Net and ensemble learning has the potential to significantly impact medical imaging, helping radiologists make better clinical decisions.