Federated Attention-Based Deep Learning Framework For Privacy-Preserving Multi-Center Knee Osteoarthritis Diagnosis From Heterogeneous Radiographic Images

Abstract

The diagnosis of Knee Osteoarthritis (KOA) using radiography is faced with issues such as different imaging conditions, unbalance in classes, low inter-center cooperation, and privacy of patients. The research presents a Distributed Attention-Based Deep Learning System that enables KOA severity classification while preserving privacy across medical facilities. In the system, image preprocessing and DenseNet121-based extraction of features are combined with the method of attention, as well as Federated Averaging (FedAvg) and Kellgren-Lawrence (KL) classification. The preprocessing of images implies activities such as resizing (224 x 224), intensivity normalization, contrast enhancement, noise reduction, region of interest extraction, augmentation, and client-based systematization is applied. The system development is based on 9,547 x-ray images, including: 5,610 images used for training, 810 images for validation, 1,630 images for internal testing, and 1,497 images for external testing. The obtained images have 3,740 grade 0 images (39.17%) and 287 grade 4 images (3.01%), which shows the significant difference across classes. The knowledge gained in federated training comes from model parameters update, as the original x-rays remain in the owners’ location.