Segmentation-Assisted Fully Convolutional Neural Network Enhances Deep Learning Performance to Identify Proliferative...

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Diabetic retinopathy (DR) progression to the proliferative stage (PDR) significantly increases vision impairment risk.
• Early detection and intervention for PDR are clinically vital.
• Current deep learning methods for DR classification can be improved.

Purpose of the Study:

• To develop and evaluate a segmentation-assisted methodology for classifying diabetic retinopathy (DR) stages, specifically PDR.
• To improve the accuracy of identifying proliferative diabetic retinopathy (PDR) using deep learning.
• To demonstrate the benefit of incorporating segmentation of retinal neovascularizations (NV) into DR classification.

Main Methods:

• Utilized the Kaggle EyePacs dataset of retinal fundus photographs.
• Employed a fully convolutional network (FCN) for segmenting retinal neovascularizations (NV).
• Trained a convolutional neural network (CNN) on combined images and probability maps for PDR classification.

Main Results:

• The segmentation-assisted classification achieved a mean accuracy of 87.71% (SD = 7.71%) on the test set.
• Segmentation-assisted PDR classification accuracy was 7.74% higher than classification alone.
• The method successfully improved the identification of the PDR stage.

Conclusions:

• Segmentation assistance significantly enhances the accuracy of identifying proliferative diabetic retinopathy (PDR).
• This approach holds potential for improving deep learning performance in medical imaging, especially with limited data.
• The methodology offers a promising tool for clinical decision support in managing diabetic retinopathy.