Glaucoma Detection Based on Joint Optic Disc and Cup Segmentation Using Dense Prediction Transformer

Glaucoma is an eye disease resulting from damage to the optic nerve. Glaucoma can cause vision loss if not treated quickly and appropriately. In this case, the observation of an ophthalmologist is needed to check it. However, an expert doctor’s observation is subjective, so it takes a long time and is inconsistent. So a computer-aided diagnostic (CAD) system was built to detect glaucoma early on, fundus image analysis automation, time efficiency through optical disc and cup segmentation, and Cup to Disc Ratio (CDR) calculations. Several previous studies have proposed using models based on Vision Transformers (ViT), Convolutional Neural Networks (CNN), and a combination of the two. However, the ViT model has problems: the number of model computations increases when the image size also increases, and the CNN-based encoder-decoder model has a large size and is slow in its calculations. Therefore, a segmentation method that utilizes transformers as encoders and convolution as decoders was chosen, namely the Dense Prediction Transformer (DPT), which can process data in parallel. This research discusses the implementation of the DPT with a case study of the optical disc and cup segmentation on fundus images using the ORIGA dataset. The result shows that DPT is outperformed slightly by the Segmenter model with mIoU at 4.9%.