Diabetic Retinal Grading Using Attention-Based Bilinear Convolutional Neural Network and Complement Cross Entropy.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Diabetic retinopathy (DR) is a significant complication of diabetes mellitus (DM), necessitating early diagnosis for effective treatment.
• Deep learning, particularly convolutional neural networks (CNNs), has shown promise in medical image analysis, including fundus lesion detection.
• Current DR image classification methods often overlook fine-grained details and suffer from imbalanced datasets, hindering accurate lesion prediction.

Purpose of the Study:

• To develop an advanced deep learning model for improved diabetic retinopathy image classification.
• To enhance the extraction of specific image features relevant to DR diagnosis.
• To address the challenges of fine-grained property recognition and data imbalance in DR datasets.

Main Methods:

• Proposed a novel non-homologous bilinear pooling convolutional neural network model.
• Integrated an attention mechanism to improve the extraction of image-specific features.
• Evaluated the model's performance against existing popular fundus image classification models.

Main Results:

• The proposed model significantly improved prediction accuracy for diabetic retinopathy classification.
• The model effectively captured fine-grained properties of diseased fundus images.
• Performance gains were achieved while maintaining computational efficiency compared to other models.

Conclusions:

• The novel CNN model with attention mechanism offers a more accurate and efficient approach to diabetic retinopathy diagnosis.
• Addressing data imbalance and focusing on fine-grained features are key to improving deep learning-based DR detection.
• This approach holds potential for enhancing early detection and management of diabetic retinopathy.