DenseLinkNet: A Deep Convolutional Architecture for Automatic Segmentation of Corneal Endothelial Cell

Corneal endothelium cell provides vital clinical information regarding the health status of the cornea, which is crucial to monitor different corneal diseases. In this regard, segmentation of the corneal endothelial cell plays a significant role in assessing different clinical parameters, like cell density, coefficient of variation, and hexagonality, that aids clinicians in diagnosing different diseases. Since the manual observation process is both time-consuming and susceptible to errors, finding an efficient approach to automate the segmentation task is imperative. In this paper, an encoder-decoder based deep neural architecture, namely DenseLinkNet, is introduced to automate the segmentation process. In the proposed segmentation architecture, DenseNet-201 is used as an encoder, and LinkNet is used as a decoder. The encoder block exploits the output of the convolutional and dense blocks, and the generated feature map gets added with the corresponding decoder block. Following the process, a segmentation head is introduced to conclude the segmentation task. In this work, the publicly available Alizarine dataset is used for experimental purpose. The proposed DenseLinkNet architecture outperforms other segmentation networks with respect to different performance metrics. The network has achieved a dice-score of 76.82% and a Jaccard-index of 62.51%.