Joint Multi-Task Deep Learning with Cross-Task Attention for Simultaneous Lesion Segmentation, Detection, and...

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Diabetic retinopathy (DR) grading typically relies on manual analysis of retinal lesions.
• Existing automated methods often analyze lesion grading or detection in isolation.
• A unified framework for simultaneous lesion analysis and DR grading is needed.

Purpose of the Study:

• To develop and evaluate a novel deep learning framework for joint lesion segmentation, detection, and DR grading.
• To assess the model's performance within and across independent datasets.
• To provide lesion-level evidence to support automated DR grading.

Main Methods:

• A hierarchical multi-task network with a ResNet-50 encoder and feature pyramid was employed.
• A cross-task attention module integrated lesion features into the grading representation.
• Homoscedastic uncertainty weighting and a three-stage schedule were used for loss combination.
• The model was trained on the DDR dataset and validated on the IDRiD dataset.

Main Results:

• Achieved a mean lesion-segmentation Dice of 0.535 and grading accuracy of 0.823 on the DDR test set.
• Demonstrated cross-dataset performance with segmentation Dice of 0.489 and grading accuracy of 0.781 on IDRiD.
• Ablation studies confirmed the importance of the cross-task attention module for grading performance.

Conclusions:

• Joint learning of lesion localization and grading provides a transparent framework with lesion-level evidence.
• Moderate performance decline was observed on the external dataset, highlighting the impact of domain shift.
• Further validation on larger, diverse cohorts is required for clinical deployment.