Fully automatic quantification of individual ganglion cells from AO-OCT volumes via weakly supervised learning
Quantitative features of individual ganglion cells (GCs) are potential paradigm changing biomarkers for improved diagnosis and treatment monitoring of GC loss in neurodegenerative diseases like glaucoma and Alzheimer's disease. The recent incorporation of adaptive optics (AO) with extremely fast and high-resolution optical coherence tomography (OCT) allows visualization of GC layer (GCL) somas in volumetric scans of the living human eye. The current standard approach for quantification - manual marking of AO-OCT volumes - is subjective, time consuming, and not practical for large scale studies. Thus, there is a need to develop an automatic technique for rapid, high throughput, and objective quantification of GC morphological properties. In this work, we present the first fully automatic method for counting and measuring GCL soma diameter in AO-OCT volumes. Aside from novelty in application, our proposed deep learningbased algorithm is novel with respect to network architecture. Also, previous deep learning OCT segmentation algorithms used pixel-level annotation masks for supervised learning. Instead in this work, we use weakly supervised training, which requires significantly less human input in curating the training set for the deep learning algorithm, as our training data is only associated with coarse-grained labels. Our automatic method achieved a high level of accuracy in counting GCL somas, which was on par with human performance yet orders of magnitude faster. Moreover, our automatic method's measure of soma diameters was in line with previous histological and in vivo semi-automatic measurement studies. These results suggest that our algorithm may eventually replace the costly and time-consuming manual marking process in future studies.
