Mixed Federated Learning Approach for Multiclass Image Segmentation
Ega Praneeth, Yeshwant Reddy, Dara Abhiram, Mukkamula Venu Gopalachari, Kiranmaie Puvulla, Salakapuri Rakesh · 2024
Image Segmentation is a computational visual task that entails partitioning a medical image into various segments, with each segment denoting a distinct object or structure of significance within the image. Nevertheless, current approaches for training segmentation models often make an impractical assumption: they presuppose that the training data for every client adheres to a uniform pattern and shares the same level of image supervision. Certainly, federated learning has risen to prominence as an innovative strategy for tackling the difficulties presented by dispersed and distant databases. It eliminates the requirement for raw data sharing by enabling numerous clients to cooperatively train a machine learning model. Federated learning aggregates the local models trained at individual data stores and leverages them to build a global model that benefits all participating entities. Based on our survey, Fed Match addresses the challenge of leveraging unlabeled data from different clients in federated learning. This is achieved by first distributing a consistency matrix among clients within the embedding space, which is used to designate helper models for the clients involved. This study aims to develop a federated learning system for medical image segmentation that will allow many medical institutions to collaborate in training a segmentation model without sharing confidential patient information. Model improvement is enabled while protecting patient privacy and sensitive patient data. In this research, an integrated label-agnostic federated learning architecture called FedMix is applied, which combines a variety of picture labels in order to segment medical pictures. It ensures that every client contributing positively to the federated model does so by leveraging all accessible labeled data. With the use of medical image datasets such as the HAM10K, UDIAT, and BUS datasets, this project aims to experiment with medical picture segmentation.