Compact Convolutional Neural Networks for Pterygium Classification using Transfer Learning

Abstract

Pterygium is one of the eye diseases that are prevalent among the workers who are frequently exposed to sunlight such as farmers, construction workers and fishermen. It is an eye disease that can be screened easily by health practitioners. However, most of the previously mentioned workers are rarely aware of the pterygium existence. Hence, health practitioners usually visit them to perform basic health checkup that includes eye disease screening. Due to many tests that need to be performed for each checkup cycle, a modern tool is required to make the screening process faster. As a result, this paper discusses a compact convolutional neural networks (CNN) approach to pterygium classification using transfer learning. Its architecture consists of only two layers of CNN and two layers of fully connected (FC) components. Rectified linear unit is used as the activation function except for the last FC layer, which utilizes Softmax function. The weights and biases are transferred from the first two layers of VGG-M architecture. Three regularization methods were experimented that include local response normalization, batch normalization and dropout. The best network configuration in term of accuracy with 0.9833 value is obtained by using local response normalization with dropout. On the other hand, the best area under the curve performance of 0.9865 is achieved by using batch normalization and dropout. Therefore, the developed compact system has managed to classify well the pterygium image taken from the frontal view.