Diabetic retinopathy detection using red lesion localization and convolutional neural networks.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Diabetic retinopathy (DR) detection is crucial for preventing vision loss.
• Automated lesion localization in retinal images aids specialists in DR diagnosis.
• Deep learning offers potential for improving DR detection efficiency and accuracy.

Purpose of the Study:

• To design a deep network patch-based lesion localization model for diabetic retinopathy.
• To reduce model complexity while enhancing performance in DR detection.
• To develop an efficient patch selection strategy for improved training.

Main Methods:

• A deep network patch-based approach was utilized for lesion localization.
• An efficient procedure involving two convolutional neural network models was designed for training patch selection.
• The model was trained on the DIARETDB1 database and tested on multiple databases, including Messidor.

Main Results:

• The model achieved an area under the receiver operating characteristic curve of 0.912 for DR screening.
• Sensitivity for DR screening reached 0.940.
• The model demonstrated competitive performance compared to state-of-the-art approaches without further adaptation.

Conclusions:

• The developed lesion localization model shows high accuracy in DR screening.
• The patch-based deep learning approach is effective in detecting early signs of diabetic retinopathy.
• This automated system can assist specialists in timely DR diagnosis and treatment, potentially preventing vision loss.