A Cross-Domain Weakly Supervised Diabetic Retinopathy Lesion Identification Method Based on Multiple Instance...

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Diabetic retinopathy (DR) detection relies on accurate lesion identification.
• Current methods require extensive fine-grained annotations and perform poorly across different institutions.
• Cross-domain application of DR detection models is a significant challenge.

Purpose of the Study:

• To develop a cross-domain, weakly supervised method for DR lesion identification.
• To reduce the reliance on detailed annotations by using coarse-grained lesion attribute labels.
• To improve the performance and generalizability of DR detection models.

Main Methods:

• Proposed a novel lesion-patch multiple instance learning (LpMIL) method for patch-level supervision.
• Developed a semantic constraint adaptation (LpSCA) method to enhance cross-domain performance.
• Created the largest fine-grained annotated dataset, EyePACS-pixel, for validation.

Main Results:

• The proposed method accurately and comprehensively identifies DR lesions.
• Achieved competitive results compared to existing detection and segmentation methods.
• Demonstrated robust performance using only coarse-grained annotations on public and custom datasets.

Conclusions:

• Weakly supervised learning with coarse labels is effective for DR lesion identification.
• The LpMIL and LpSCA methods offer a viable alternative to fine-grained annotation-dependent approaches.
• This work advances the clinical application of automatic DR detection by improving cross-domain adaptability.