An Explainable Federated Learning Approach for Pulmonary Nodules Diagnosis

With over two million new cases of lung cancer estimated globally, lung cancer remains the leading cause of death every year. Computer-aided diagnostic systems have revolutionized lung cancer diagnosis using deep learning and computed tomography imaging. However, low model reliability and small, low-quality, and inconsistent datasets persist. Therefore, this reinforces the call for collaborative systems to enable a venue for getting better, more generalizable results while at the same time ensuring patient data privacy. In particular, this paper presents a new kind of decentralized yet explainable model that would enhance lung cancer diagnosis. The proposed approach adopts a federated learning architecture based on MobileNetv2 with 2D-CNN, using four different datasets from Kaggle to evaluate the model. Moreover, it was tested using an external dataset to assess its generalization capability, with CAM-GRAD employed to improve its trustworthiness and comprehensibility. The model successfully identified all negative cases with a 0% false positive rate and a 2.46% false negative rate, resulting in an accuracy of 98.76%. The model reached the same performance when tested on an unseen dataset, achieving identical results.