Cross-Institutional Breast Tumor Detection via Federated Learning and Convolutional Neural Networks
Shiva Mehta, Rajat Saini · 2024
Among all cancer diseases, breast cancer remains one of the leading causes of cancer-related death, thus the need for enhanced methods for early identification. This study seeks to determine the applicability of federated learning techniques in detecting breast tumors using a CNN among five clients. Every individual, each client, works on a model that can be shared while at the same time ensuring that data cannot be compromised. The dataset consisted of 3,600 breast tumor photos that were divided into five categories: benign, malignant, cysts, fibroadenomas, and other nonmalignant tumors. Regarding the performance measures, the federated model achieved a 92% accuracy rate, 91% precision rate, 90% recall rate, and an F1 score of 90%. The outcome of this model is almost to a centralized model’s performance; it has an accuracy of 93 percent, with 92 percent precision, a recall of 91 percent, and an F1 score of 91 percent. The multiple advantages demonstrated by the federated learning technique include better data protection, the ability to utilize data from various sources belonging to different customers, and the reduction in the need for extensive centralized computations. However, we have identified a few challenges in this process, including the fact that there is often a need to over-communicate the process, the implementation process is complex, and the variation of performance due to data differences.