Federated Learning with Global Model Hint for Medical Image Object Detection

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Computer Vision

Background:

• Medical image object detection requires extensive data, often limited by annotation costs and privacy concerns.
• Federated learning (FL) offers a solution for data privacy but suffers from amplified feature drift in medical image object detection.

Purpose of the Study:

• To propose FedMHDet (Model Hint Federated Learning Detection Model), a novel FL framework to mitigate feature drift in medical image object detection.
• To enhance the performance of FL-based medical image object detection models.

Main Methods:

• FedMHDet leverages multi-scale feature consistency as a global model hint during training.
• This hint guides client models to reduce feature drift and improve convergence.
• The framework was tested on pulmonary lesion and brain tumor detection tasks.

Main Results:

• FedMHDet demonstrated favorable overall performance in pulmonary lesion and brain tumor detection.
• It improved average AP by 1.05 and 0.19, and average sensitivity by 1.10 and 0.43 respectively, compared to the strongest baselines.
• In-depth analyses support the practical utility of the proposed method.

Conclusions:

• FedMHDet effectively addresses the feature drift problem in federated learning for medical image object detection.
• The proposed model hint strategy enhances detection accuracy and sensitivity.
• FedMHDet offers a promising approach for privacy-preserving medical image analysis.