Support vector machine and deep-learning object detection for localisation of hard exudates

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Hard exudates are key indicators in diabetic retinopathy, necessitating accurate and early detection.
• Current detection methods can be improved for speed and reliability.

Purpose of the Study:

• To develop a novel, automated method for identifying and localizing hard exudates in retinal images.
• To enhance the efficiency and accuracy of hard exudate detection systems.

Main Methods:

• A hybrid approach combining a Support Vector Machine (SVM) classifier for rapid pre-scanning with a Faster Region-based Convolutional Neural Network (Faster R-CNN) object detector.
• Utilized ResNet-50 for feature extraction during the pre-scanning phase to filter out exudate-free images.
• Detailed analysis of remaining images using the Faster R-CNN for precise exudate localization.

Main Results:

• The SVM pre-scanning reduced the false positive rate by 29.7% when evaluating individual exudates, with a marginal 16.2% increase in false negatives.
• A 50% reduction in false positive rate was achieved when evaluating entire images, without increasing false negatives.
• The combined method demonstrated potential for simultaneous improvement in speed and accuracy, particularly with limited training data.

Conclusions:

• Pre-scanning retinal images with an SVM before applying a Faster R-CNN detector can significantly improve hard exudate detection.
• This novel method offers a promising solution for faster and more reliable automated detection of hard exudates in diabetic retinopathy.
• The approach is especially beneficial in scenarios with limited training data, enhancing diagnostic capabilities.