Publication: Optic disc detection and segmentation from fundus image using convolutional neural network and localize active contour
Loading...
Date
2023-09-01
Authors
Poh Chyong Yi
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
The detection and segmentation of optic disc (OD) from fundus images is a crucial task in the development of automated fundus diagnosis systems. However, this task presents several challenges due to variations in image quality, the presence of noise, and differences in the appearance of the OD across individuals. In this research, a method of OD detection based on deep learning approach is proposed. The method uses Faster Region-based Convolutional Neural Network (Faster RCNN) framework with pretrained Convolutional Neural Network (CNN) as its backbone feature extractor. The performance of the OD detection system was evaluated using different pretrained CNNs and image pre-processing steps. The types of the pretrained CNNs used include AlexNet, VGG-19 and ResNet-50. The experiment results show that Faster RCNN with Residual network 50 (ResNet-50) pretrained CNN, using image pre-processing based on image enhancement using Contrast-limited Adaptive Histogram Equalization (CLAHE), image augmentation and CIEXYZ colour format produced the best overall results. It achieved precision, sensitivity and miss rate of 0.9211, 0.9211 and 0.0789 respectively. The result of OD detection is a bounding box that marks the region of OD in the fundus image. This box is then cropped out and processed in the OD segmentation stage. In this research, the performance of four image segmentation methods, namely Superpixel-based Fast Fuzzy C-Means Clustering (SFFCM), K-means clustering, active contour and localised region based active contour were evaluated for OD segmentation. The results show that localized region-based active contour achieved the best performance. The performance of OD segmentation based on localized region-based active contour was further improved by adding an image pre-processing stage. The pre-processing steps include image multilevel thresholding based on Fractional Order Darwinian Particle Swarm Optimization (FODPSO) and morphological operation. Evaluation using the Drishti-GS fundus image dataset shows that the proposed method outperformed other methods of OD segmentation. It achieved the highest Dice and Jaccard coefficients of 0.9550 and 0.9134 respectively.