Pterygium Tissues Segmentation using Densely Connected DeepLab

One of the eye conditions that are prevalent among the fishermen is pterygium. It happens when abnormal fibrovascular tissues grow towards the pupil region. It can cause vision problems if the tissues fully obstruct light from coming into the retina. Thus, the disease must be detected as early as possible so that the mitigation steps can be administered before it is too late. However, most of the fishermen are not aware of the pterygium existence. Therefore, an online screening tool should be developed to help them doing screening by themselves just by using a mobile phone. Taking inspiration from this need, an automatic pterygium tissue segmentation is proposed using a deep learning approach. The goal is to extract the right size of the pterygium infected tissues so that the users are aware of their disease severity level. The proposed method uses DeepLab V1 and DeepLab V2 as the foundation, in which densely connected layers are integrated to improve the segmentation performance. A set of four feedforward layers is added to the existing DeepLab architectures with the aim of reducing the zero gradient diminishing problem. Both DeepLab V1 and V2 have recorded performance improvement in terms of mean accuracy (acc) and mean intersection over union (IoU). Densely connected DeepLab V2 obtains the best testing results with acc and IoU of 0.9202 and 0.8381, respectively. Hence, the improved segmentation method is suitable to be implemented for the mobile pterygium screening system. The segmentation performance can be further enhanced by adding more parallel layers or convolution-based down-pooling operation.