D-TNet: a hybrid Dense Net-transformer model for robust diabetic retinopathy detection

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Diabetic Retinopathy (DR) is a leading cause of blindness and a serious diabetes complication.
• Early DR detection and severity assessment are crucial for preventing vision loss.
• Manual DR diagnosis from retinal images is time-consuming and subjective.

Purpose of the Study:

• To develop an advanced AI model for accurate Diabetic Retinopathy severity grading.
• To overcome limitations of existing AI methods in DR classification performance and generalizability.

Main Methods:

• A hybrid Deep Learning (DL) model, D-TNet, was developed, combining DenseNet121 and a Transformer architecture.
• The model was trained and tested on retinal images from APTOS2019 and Messidor-2 datasets.
• D-TNet identifies key DR indicators like microaneurysms, hemorrhages, and neovascularization.

Main Results:

• D-TNet achieved 97% accuracy on the APTOS2019 dataset and 86% accuracy on the Messidor-2 dataset.
• High F1-scores (0.94 and 0.79) and kappa scores (0.93 and 0.80) were obtained on both datasets.
• The model demonstrated robust and balanced classification across all five DR severity stages.

Conclusions:

• The D-TNet model offers a stable and balanced approach for DR severity grading, outperforming traditional AI methods.
• This hybrid DL strategy has the potential to enhance diabetic eye care, especially in resource-limited settings.
• Future work includes multimodal data integration and domain adaptation for real-world deployment.