Year: 2024
Type: article
Abstract: The widespread application of deep learning in the medical field has been increasingly prominent. The utilization of deep learning for the recognition and classification of tuberculosis (TB) in CT medical images has become a popular research topic. However, federated learning exhibits unique advantages. Besides accomplishing tasks that deep learning can perform, it also has the ability to protect patients' privacy to a certain extent. This characteristic makes it more readily acceptable. Nevertheless, in real-world scenarios, data heterogeneity among federated learning participants can lead to model bias, hindering the achievement of desired results. To address this challenge, this study proposes a method that employs client data distribution information clustering and a loss function suppression term. This approach demonstrates excellent performance in both data-heterogeneous and non-data-heterogeneous scenarios. Experimental validation on a TB CT image dataset confirms its effectiveness. Compared to traditional federated learning baseline methods such as Federated Averaging (FedAvg), our method achieves an improvement of up to 11.54% in model accuracy, with faster convergence speed. Moreover, it exhibits greater stability in model accuracy when subjected to communication cost constraints.
Authors: Zhihan Yang, Xiaogang Wang
Institution: Chongqing University of Science and Technology
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: diamond
OpenAlex ID: W4409052501