A Study of Convolution Neural Network Based Cataract Detection with Image Segmentation

Timely and precise cataract detection is crucial to managing the risk and preventing blindness for cataract's patients. This paper proposed a framework for automatic cataract detection consisting of the K-Means clustering-based segmentation (KMSeg) and Convolutional Neural Network (CNN). At first, data pre-processing was performed. Then, KMSeg is responsible for characterizing the input images into a subgroup of color. Lastly, three CNN were employed based on DCNN, ResNet18, and ResNet50 backbones for feature learning and classification task. An extensive study was examined on Fundus and Front Eye datasets with numerous experimental settings. The result shows that the proposed KMSeg-CNN is able to maintain accuracy yet provides a faster training and testing execution time across the dataset.