Improving detection and classification of diabetic retinopathy using CUDA and Mask RCNN

Diabetic retinopathy (DR) is an eye disease caused by diabetes and can progress to certain degrees. Because DR’s the final stage can cause blindness, early detection is crucial to prevent visual disturbances. With the development of GPU technology, image classification and object detection can be done faster. Particularly on medical images, these processes play an important role in disease detection. In this work, we improved our previous work to detect diabetic retinopathy using Faster RCNN and attention layer. In the detection phase, firstly non-used area of DR image was extracted using compute unified device architecture with gradient-based edge detection method. Then Mask RCNN was used instead of faster region-based convolutional neural networks (Faster RCNN) to detect lesion areas more successful. With the proposed method, more successful results were obtained than the our previous work in DenseNet, MobileNet and ResNet networks. In addition, more successful results were obtained than other works in the literature in ACC and AUC metrics obtained by using VGG19.