Retinal Layer and Fluid Segmentation with Transformer Based Architecture.

Retinal layer and fluid segmentation is a critical task in assisting doctors to diagnose retinal diseases. Manual segmentation by experts provides the highest accuracy, but it is time-consuming and inconsistent if segmented by different experts. Deep learning algorithms(e.g. Convolutional Neural Network(CNN)) have provided a faster way to perform segmentation through a computer-aided diagnosis system. Nevertheless, CNN has limitations, such as a limited receptive field and loss of details. In this project, we propose a transformer-based architecture to segment the retinal layer and fluid from retinal images. The architecture is based on Vision Transformer (ViT) and modified to improve performance. The transformer has been trained on a set of training retinal images and evaluated on a separate set of testing retinal images. The transformer-based architecture demonstrated a 0.01 improvement in average dice coefficient compared to the Unet architecture for fluid and layer segmentation. The Transformer-based architecture is better suited for deployment in commercial portable Optical Coherence Tomography (OCT) devices due to significantly faster inference speed. The proposed model is at most 4 times higher than that of the CNN family models. This makes it an ideal choice for resource-constrained environments where computational resources are limited.