DR-ConvNeXt: DR classification method for reconstructing ConvNeXt model structure.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Diabetic retinopathy (DR) is a significant complication of diabetes, contributing to vision loss in the working-age population.
• Classifying DR from retinal images is challenging due to complex lesion distributions and variability.
• Accurate DR detection is crucial for timely intervention and preventing blindness.

Purpose of the Study:

• To develop an automated method, DR-ConvNeXt, for precise classification of diabetic retinopathy lesion types.
• To enhance the accuracy of diabetic retinopathy diagnosis through advanced deep learning techniques.

Main Methods:

• The DR-ConvNeXt model utilizes a dual-branch addition convolution structure.
• Increased stacked ConvNeXt Block convolution layers were incorporated.
• A unique primary-auxiliary loss function was introduced to improve classification performance.

Main Results:

• On the APTOS dataset, DR-ConvNeXt achieved 91.8% accuracy, 81.6% sensitivity, and 97.9% specificity.
• On the Messidor-2 dataset, the model obtained 83.6% accuracy, 74.0% sensitivity, and 94.6% specificity.
• The model demonstrated superior performance across all evaluation metrics on both datasets.

Conclusions:

• The DR-ConvNeXt model exhibits significant advantages in classifying diabetic retinopathy.
• The proposed method offers a promising approach for accurate and automated DR diagnosis.
• The enhanced classification performance highlights the potential of DR-ConvNeXt in clinical settings.