Application of a Twin Parametric Support Vector Machine and Deep Learning Techniques for Abnormal Retinopathy Detection from Fundus Photographs

Abstract

In today’s world, abnormal retinopathy detection from fundus photographs plays a major role in a non-destructive ocular diagnosis. There are some applications of machine learning in ophthalmic imaging that help in the prescreening of ophthalmologists. This research proposes application of of a twin parametric support vector machine and deep learning techniques for the classification of normal and abnormal retinopathy from fundus photographs. The data is an open source of fundus photographs from Shanggong Medical Technology (SMT) Co., Ltd. by Kaggle.com. The procedure started with the image processing parts. All photographs were processed using 5 image processing techniques, which were ridge detection, ridge detection by Robert edge detection and hysteresis thresholding method, ridge detection by Sobel edge detection and hysteresis thresholding method, adaptive thresholding, and color contrast enhancement. The non-processing and processed images were resized, and then features were extracted by 5 algorithms of a convolutional neural network, ResNet50, InceptionV3, MobilenetV2, VGG19, and DenseNet201. A classical support vector machine method and a twin parametric support vector machine method were used for the classification. It was found that the classical support vector machine with the  MobilenetV2 feature extraction algorithm provided the model with the best accuracy, 92.98%, for 1344.0234 seconds of computational time. On the other hand, the classification by the twin parametric support vector machine combined with all five feature selection algorithms performed with a bit lower accuracy but much faster computational time. The accuracy of the models created by the twin parametric support vector machine ranged from 84.74% to 87.64%. The model by the twin parametric support vector machine combined MobilenetV2 was created fastest by 111.5658 seconds only, with 86.98% accuracy. Overall, the twin parametric support vector machine offered the model with more or less 6% lower accuracy but 10 times faster than the classical one.