Multi-Model Domain Adaptation for Diabetic Retinopathy Classification.

Area of Science:

• Ophthalmology
• Computer Science
• Medical Imaging

Background:

• Diabetic retinopathy (DR) is a leading cause of blindness, necessitating efficient screening methods.
• Current DR screening relies on experienced ophthalmologists, which is time-consuming and prone to misdiagnosis.
• Deep learning models show promise for DR diagnosis but require extensive labeled datasets, which are costly to obtain.

Purpose of the Study:

• To develop a novel unsupervised domain adaptation method for diabetic retinopathy classification.
• To address the challenge of limited labeled data in medical image analysis, specifically for DR detection.
• To enable DR classification using unlabeled retinal images by leveraging information from multiple source models without direct data access.

Main Methods:

• A multi-model domain adaptation (MMDA) technique was developed for unsupervised DR classification.
• The method utilizes discriminative information from multiple source models without accessing their data.
• A novel weighting mechanism was integrated to assess source domain importance, coupled with a weighted pseudo-labeling strategy.

Main Results:

• The MMDA method demonstrated competitive performance against state-of-the-art approaches on the APTOS 2019 dataset.
• Experiments were conducted using four source datasets (DDR, IDRiD, Messidor, Messidor-2) targeting the APTOS 2019 dataset.
• The proposed approach achieved effective DR classification in an unsupervised, unlabeled target domain.

Conclusions:

• The study presents a novel domain adaptation solution for medical image analysis, particularly for DR detection when source data is unavailable.
• MMDA offers a viable alternative to traditional deep learning models that require labeled datasets.
• This research contributes a new approach to unsupervised DR classification, enhancing the potential for automated screening.