Imam Syahrizal, Izzun Nafis Ibadik, Pandega Abyan Zumarsyah
[email protected], [email protected], [email protected]
Department of Electrical Engineering and Information Technology, Universitas Gadjah Mada
Fundus image analysis is valuable for detecting various eye diseases such as glaucoma and retinopathy. This study explores several methods to achieve optimal segmentation results: region segmentation with GrowCut, dynamic thresholding and GrabCut, Random Forest, and UNet-MobileNetV2. These methods were evaluated using three datasets: ORIGA, DRIONS-DB, and RIM-ONE v3.
The UNet-MobileNetV2 model delivered excellent results across all three datasets. The Random Forest model also performed segmentation on all datasets but did not yield satisfactory results on the RIM-ONE v3 dataset. In contrast, methods 1 and 2 produced good results only when applied to the local dataset. Without parameter adjustments, these methods struggled to match the performance of the UNet-MobileNetV2 model.
This program segments the optic disc in fundus images using various datasets and methods.
- Dataset 1 (ORIGA: source): Contains 100 fundus images with a resolution of 3072x2048 pixels. Ground truth is provided as .mat files containing optic disc and optic cup region information.
- Dataset 2 (DRIONS-DB): Contains 110 fundus images with a resolution of 600x400 pixels. 23% of the images are from glaucoma patients, while the remaining 77% are from ocular hypertension patients.
- Dataset 3 (RIM-ONE r3): Contains 159 fundus images. These images are represented as horizontally arranged stereo images with annotations for the left optic disc.
- Method 1: Preprocessing => Blood Vessel Removal => Circular Hough Transform => GrowCut
- Method 2: Four interrelated methods utilizing Thresholding, Grabcut, and Fit-Ellipse
- Method 3: Machine Learning with a Random Forest model for segmentation
- Method 4: Deep Learning based on UNet architecture with a MobileNetV2 model implemented using PyTorch