Collaborative Deep Learning for Privacy Preserving Diabetic Retinopathy Detection.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Diabetic Retinopathy (DR) detection using Convolutional Neural Networks (CNNs) shows promise but requires large datasets.
• Privacy concerns and data access limitations hinder the development of robust DR detection models.
• Existing methods struggle with data scarcity and privacy issues in medical imaging.

Purpose of the Study:

• To propose a novel collaborative learning framework for training CNN models for DR detection.
• To address data privacy concerns by enabling distributed learning without data sharing.
• To improve the accuracy of DR detection by leveraging multiple, distributed datasets.

Main Methods:

• Developed a privacy-preserving collaborative learning approach for training CNNs in a distributed environment.
• Implemented a transfer learning strategy where models are iteratively retrained across datasets in ascending order of performance.
• Trained and tested CNN models on five different retina image datasets.

Main Results:

• Achieved a DR detection accuracy of 93.5% using the proposed collaborative learning method.
• Outperformed accuracies from merged datasets (84%) and individual datasets (ranging from 73% to 85%).
• Demonstrated the effectiveness of the ascending chain transfer learning approach.

Conclusions:

• The proposed collaborative learning method significantly improves DR detection accuracy while preserving data privacy.
• This approach offers a viable solution for training deep learning models on sensitive medical data.
• The findings highlight the potential of distributed learning in advancing medical image analysis.