Pterygium Classification Using Deep Patch Region-based Anterior Segment Photographed Images

Early pterygium screening is crucial to avoid blurred vision caused by cornea and pupil encroachment. However, medical assessment and conventional screening could be laborious and time-consuming to be implemented. This constraint seeks an advanced yet efficient automated pterygium screening to assist the current diagnostic method. Patch region-based anterior segment photographed images (ASPIs) focus the feature on a particular region of the pterygium growth. This work addresses the data limitation on deep neural network (DNN) processing with large-scale data requirements. It presents an automated pterygium classification of patch region-based ASPI using our previous re-establish network, "VggNet16-wbn", the VggNet16, with the addition of batch normalisation layer after each convolutional layer. During an image pre-processing step, the pterygium and nonpterygium tissue are extracted from ASPI, followed by the generation of a single and three-by-three image patch region-based on the size of the 85x85 dataset. Data preparation with 10-fold cross-validation has been conducted to ensure the data are well generalised to minimise the probability of underfitting and overfitting problems. The proposed experimental work has successfully classified the pterygium tissue with more than 99% accuracy, sensitivity, specificity, and precision using appropriate hyperparameters values. This work could be used as a baseline framework for pterygium classification using limited data processing.