Abstract
Purpose :
In vivo confocal microscopy (IVCM) is non-invasive, reproducible, and inexpensive diagnostic tool for corneal diseases, which also helps to detect neurological and metabolic diseases. However widespread and effortless image acquisition with IVCM creates also a serious workload of image processing and analysis for clinicians. Deep learning algorithms are possible solutions for this heavy workload. In our study, we have produced a novel deep learning algorithm based on generative adversarial network (GAN) and we compare its accuracy for automatically segmentation of subbasal nerve plexus in IVCM images with ophthalmology experts and a convolutional neural network(U-Net) based method.
Methods :
We have collected IVCM images from various patient groups and anonymized patient information for segmentation procedures. Three graders, U-Net based conventional algorithm and our GAN-based nerve segmentation system traced nerve plexus for each IVCM images. Results for GAN and U-Net based nerve segmentation methods in IVCM images compared with the graders and analyzed with Pearson's r correlation, Bland-Altman analysis, and receiver operating characteristics (ROC) curves. Lastly, different types of noises applied on IVCM images and segmentation performance of GAN-based and U-Net based methods observed under these different noise types.
Results :
The GAN-based algorithm demonstrated similar correlation and Bland-Altman analysis results compared to U-Net-based algorithm. When ROC curves applied to both methods, the GAN-based segmentation method showed significantly higher sensitivity and higher specificity compared to U-Net based algorithm for corneal nerve segmentation in IVCM images (p<.001). Lastly, the performance of the U-Net-based algorithm deteriorates significantly with real world noise simulation, especially in speckle type noise, compared to GAN-based algorithm (p<.001).
Conclusions :
This study is the first study for application of GAN-based algorithms on IVCM images. The GAN-based algorithm demonstrated higher accuracy compared to the experts and the U-Net based algorithm for corneal nerve segmentation in IVCM images. The GAN-based algorithm is more reliable than the U-Net based method in IVCM images with different types of noise. In the near future, the GAN-based segmentation method could be used as a facilitative diagnostic tool in ophthalmology clinics.
This is a 2021 ARVO Annual Meeting abstract.