|
Corneal nerve fibers (CNF) usually exhibit changes associated with ocular and systemic diseases. Thus, in clinic, CNF is imaged by corneal confocal microscopy (CCM) for ocular disease diagnosis. To obtain an objective and accurate diagnosis, CNF needs to be segmented from CCM image and further its centerline is extracted to pave the wave for producing quantitative diagnostic markers. It is reasonable to state that the performance of CNF segmentation and centerline extraction places a big impact on the disease diagnosis. Therefore, in this study, we aim to improve the CNF segmentation and centerline extraction. CNF segmentation algorithm is developed based on UNet++, which is modified by adding skip connection to be more applicable to multi-scale information. Following CNF segmentation, a new CNF centerline extraction algorithm is developed based on neighborhood statistics. Compared with the traditional thinning algorithm and the skeleton extraction algorithm, our method yields a more consistent and smoother central line extraction by reducing the effect of imperfect segmentation and noise which is hard to avoid in deep learning segmentation. The proposed segmentation method is evaluated on an open dataset by comparing with the conventional UNet, UNet++ and UNet3+. The proposed centerline extraction method is evaluated on the same image dataset by comparing with the traditional thinning algorithm and the skeleton extraction algorithm. The results show that our method can outperform the conventional UNet++ in terms of Acc (accuracy), TPR (True Positive Rate), TNR (True Negative Rate), Dice and FDR (False Discovery Rate). The centerline extraction method can extract the centerline with less errors.
|
