WWFedCBMIR: World-Wide Federated Content-Based Medical Image Retrieval

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Pathology

Background:

• Content-based medical image retrieval (CBMIR) aids pathologists by finding similar cases.
• Training CBMIR models requires diverse whole-slide images (WSIs), which are difficult to obtain due to data sharing regulations.
• Federated learning (FL) offers a solution for decentralized model training without compromising data privacy.

Purpose of the Study:

• To propose a federated content-based medical image retrieval (FedCBMIR) tool.
• To address the challenge of acquiring diverse medical datasets for training CBMIR models.
• To improve the performance and efficiency of CBMIR through federated learning.

Main Methods:

• Developed an unsupervised feature extractor (FE) for CBMIR.
• Implemented a federated learning framework (FedCBMIR) to distribute the FE to collaborative centers for training.
• Evaluated FedCBMIR using two experiments with breast cancer datasets (BreaKHis, Camelyon17) across different clients and magnifications.

Main Results:

• FedCBMIR significantly improved F1 scores: from 96% to 98.1% (Camelyon17) and 95% to 98.4% (BreaKHis).
• Reduced training time by 11.44 hours in the initial experiments.
• Achieved high F1 scores (94-98%) with a generalized model in the BreaKHis experiment, completing training 25.53 hours faster.

Conclusions:

• FedCBMIR effectively trains CBMIR models using federated learning without centralizing sensitive medical data.
• The proposed method enhances retrieval accuracy and substantially decreases model training duration.
• FedCBMIR demonstrates a promising approach for developing robust CBMIR tools in healthcare settings.