Diabetic Retinopathy (DR) is a fast-progressing disease, often resulting in blindness, early diagnosis being crucial to prevent further damage. Eye Fundus Images (EFI) can be analyzed to detect lesions and the degree of DR. Automated detection of individual lesions helps visualizing the lesions, characterizing their location, size and severity, and also detecting the degree of DR [1]. Deep learning is state-of-the-art in segmentation procedures. Aim: Evaluate state-of-the-art, deep learning-based, segmentation of EFI.
IDRID Diabetic Retinopathy dataset with 55 train and 28 test Eye Fundus Images (EFI), together with corresponding groundtruth label masks; DeepLabV3 vs Fully Convolutional Network (FCN): segmentation using deep convolution neural networks (DCNN) trained to recognize the lesions; DCNN is used to recognize lesions in test dataset; Metrics: (A) = Global Accuracy and Iou; (B) Per-lesion accuracy (OD=Optic Disk, HA=Haemorrhages, HE=Hard Exudates, SE=Soft Exudates, MA=MicroAneurisms, BK=bkground);
Training Runtime: DeepLabV3 19 minutes with 1 GPU; FCN 635 minutes with 1 GPU; Accuracy (A) DeepLabV3 (acc 82.2%, IoU 32%), FCN (88.5%, IoU 38%); (B) DeepLabV3 (OD 95.3%, SE 84.1%, HA 65.3%, HE 95.2%, MA 83.9%, BK 80%), FCN (OD 95.3%, SE 62.1%, HA 58.1%, HE 80.2%, MA 63%, BK 89%).
Results show that DCNN approaches achieve relatively high accuracy but low IoU, and DeepLabV3 achieves better accuracy over individual lesions. Future work: the approaches need further developments to improve IoU, our focus of future work on the subject.