Purchase this article with an account.
Jonathan D Oakley, Daniel B Russakoff, Rachel Weinberg, Megan McCarron, Samuel Brill, Stuti Misra, Charles N McGhee, Joseph Mankowski; Automated Analysis of In Vivo Confocal Microscopy Corneal Images Using Deep Learning. Invest. Ophthalmol. Vis. Sci. 2018;59(9):1799.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
In vivo confocal microscopy (IVCM) allows for non-invasive, cellular level imaging of the cornea. This facilitates detailed corneal nerve assessment offering objective biomarkers that have clinical relevance in a number of diseases. With variability in IVCM image quality, automated algorithms remain less accurate than expert assessment by manual tracing. In this study, we address such deficiencies using a novel deep learning approach in automated assessment of variant quality images.
IVCM images were acquired from 29 macaques using the Heidelberg HRT-3 and Rostock Cornea Module. Macaque corneal images were of lower quality than human images because of the smaller radius of curvature and the need to manually position animals for confocal scanning under general anesthesia. To maintain a strong correlation to expert readers we replaced our previously reported detection mechanism with a deep convolutional neural network (CNN), trained to recognize pixels of nerves or background (Fig. 1). The same pre- and post-processing components were kept for their effective noise reduction and creation of continuous nerve images, respectively. For training, 40 images were used. 46 separate images were traced by 3 experts using ImageJ (NeuronJ plugin) and used for testing. Total nerve density was compared across readers.
Algorithm agreement was good to excellent relative to all readers (Pearson coefficients of 0.80, 0.86, 0.89), similar to that between readers (0.83, 0.86, 0.91). Bland-Altman plots demonstrated inter-observer reproducibility showing tight limits of agreement (LOA) across the dynamic range of densities (Fig. 2). This is a marked improvement over our previous automated analysis method.
Deep learning based approaches to object recognition and analysis are state of the art. Book-ended by effective pre- and post-processing, we have advanced our previous automated analysis of IVCM images. With good accuracy and sensitivity in challenging data, this approach also facilitates the evaluation of higher order morphological features of nerves such as tortuosity and branching, adding additional value to IVCM analyses.
This is an abstract that was submitted for the 2018 ARVO Annual Meeting, held in Honolulu, Hawaii, April 29 - May 3, 2018.
Fig. 1. Example input image and output prediction class probabilities from the CNN.
Fig. 2. Bland Altman (LOA are 0.0051 and -0.0033) and scatter plot of total nerve fiber density (algorithm vs average reader).
This PDF is available to Subscribers Only