Detection of Corneal Ulcer Using a Genetic Algorithm-Based Image Selection and Residual Neural Network

Area of Science:

• Ophthalmology
• Computer Science
• Medical Imaging

Background:

• Corneal ulceration is a severe eye condition leading to permanent vision loss.
• Current diagnostic methods for corneal ulcers are limited.
• Deep neural networks (DNNs) show promise for disease classification but require extensive data.

Purpose of the Study:

• To enhance the classification performance of pre-trained DNNs for corneal ulcer detection using limited datasets.
• To address the issue of redundant features in deep learning models applied to medical image analysis.

Main Methods:

• Utilized transfer learning with pre-trained DNNs (e.g., ResNet) on a small-scale corneal ulcer dataset.
• Developed a novel technique employing a genetic algorithm (GA) to systematically select and remove redundant features within DNN layers.
• Tested the GA-enhanced ResNet model against classical approaches.

Main Results:

• The proposed GA-based feature selection method significantly improved the classification performance of ResNet for corneal ulcers.
• The technique effectively reduced redundancy in DNN layers, leading to better accuracy with limited data.
• Achieved superior results compared to traditional classification methods.

Conclusions:

• The GA-enhanced DNN approach offers a promising solution for accurate corneal ulcer detection, particularly when dealing with small datasets.
• This method provides a more efficient and effective way to leverage deep learning in ophthalmology.
• The systematic removal of redundant features is crucial for optimizing DNN performance in medical applications.