CABNet: Category Attention Block for Imbalanced Diabetic Retinopathy Grading

Area of Science:

• Ophthalmology
• Medical Imaging
• Computer Science

Background:

• Diabetic Retinopathy (DR) grading is complex due to subtle lesions and imbalanced datasets.
• Traditional Convolutional Neural Networks (CNNs) struggle with fine-grained DR feature identification.

Purpose of the Study:

• To develop an attention-based deep learning model for improved Diabetic Retinopathy grading.
• To address challenges of intra-class variations, small lesions, and imbalanced data distributions in DR grading.

Main Methods:

• Proposed a novel Category Attention Block (CAB) to address imbalanced DR data distributions by focusing on discriminative region-wise features.
• Introduced a Global Attention Block (GAB) to capture detailed, class-agnostic features of small lesions in fundus images.
• Integrated CAB and GAB into a backbone network to create CABNet for end-to-end DR grading.

Main Results:

• CABNet demonstrated significant performance improvements on existing deep learning architectures.
• The model achieved state-of-the-art results in Diabetic Retinopathy grading across three public datasets.
• CABNet requires few additional parameters while enhancing performance.

Conclusions:

• The proposed attention modules (CAB and GAB) effectively enhance DR grading by addressing data imbalance and small lesion detection.
• CABNet offers an efficient and effective solution for fine-grained Diabetic Retinopathy grading.
• The developed attention blocks are versatile and can be integrated with various backbone networks.