Privacy Preserving Federated Learning for Lung Cancer Classification

Lung cancer is characterized by high mortality and incidence rates, making it one of the most prevalent cancers globally. Early detection significantly improves the chances of survival for individuals affected by this disease. The histopathological diagnosis is a crucial factor in determining the specific type of cancer. In recent years, there has been a significant increase in novel computer-aided diagnostic techniques utilizing deep learning algorithms for the early detection of lung cancer. However, sharing sensitive patient data is significantly restricted by regulations such as HIPAA and GDPR, primarily due to privacy concerns. Given the current constraints, institutions face challenges in effectively exchanging information to enhance the accuracy of lung cancer classification. In order to address the issue of privacy in lung cancer classification, we propose a federated learning approach. This methodology involves employing local models with an Inception-v3 backbone to carry out the classification of histopathological images of lung cancer & updating the global model based on the local weights. These images have been obtained from the LC25000 dataset. The lung cancer images from the LC25000 dataset were analyzed, which consisted of three distinct classes. Each class contained a total of 5000 images. The applied model has achieved a classification accuracy of 99.867% in categorizing lung cancer images into three distinct classes. The performance of the proposed framework has demonstrated superiority over other existing methodologies. Furthermore, this solution effectively addresses the privacy concerns associated with the sharing of medical data among different institutions.