Improving corneal nerve segmentation using tolerance Dice loss function

Abstract In vivo confocal microscopy is a technique that allows to acquire images of the corneal layers in a rapid and noninvasive way. Analysis of sub-basal nerve allows obtaining important clinical information regarding the eye and the human body’s health. To obtain that information, it is necessary to correctly identify and trace the nerve fibers. Manual analysis is time-consuming and subjective. Numerous automatic algorithms have been proposed to overcome these problems, but none have been included in clinical practice yet. In this work, we take advantage of deep learning techniques. We present an analysis of the performances obtained through UNet (baseline) to which various architectural solutions have been added to boost performance. The variation of the tracing results is also analyzed according to the use of different loss functions, one of which is introduced here: It considers a tolerance margin (Dice with tolerance). The investigated configurations have been shown to be capable of improving the tracing of corneal nerve fibers. The model with attention modules and atrous-spatial pyramid pooling modules showed the greatest improvement compared to the baseline, increasing in the evaluation score from 86.51 to 90.21%. Furthermore, the proposed loss function further increases the results (achieving 92.44%).