Semantic segmentation of gonio-photographs via adaptive ROI localisation and uncertainty estimation.

OBJECTIVE: To develop and test a deep learning (DL) model for semantic segmentation of anatomical layers of the anterior chamber angle (ACA) in digital gonio-photographs. METHODS AND ANALYSIS: We used a pilot dataset of 274 ACA sector images, annotated by expert ophthalmologists to delineate five anatomical layers: iris root, ciliary body band, scleral spur, trabecular meshwork and cornea. Narrow depth-of-field and peripheral vignetting prevented clinicians from annotating part of each image with sufficient confidence, introducing a degree of subjectivity and features correlation in the ground truth. To overcome these limitations, we present a DL model, designed and trained to perform two tasks simultaneously: (1) maximise the segmentation accuracy within the annotated region of each frame and (2) identify a region of interest (ROI) based on local image informativeness. Moreover, our calibrated model provides results interpretability returning pixel-wise classification uncertainty through Monte Carlo dropout. RESULTS: The model was trained and validated in a 5-fold cross-validation experiment on ~90% of available data, achieving ~91% average segmentation accuracy within the annotated part of each ground truth image of the hold-out test set. An appropriate ROI was successfully identified in all test frames. The uncertainty estimation module located correctly inaccuracies and errors of segmentation outputs. CONCLUSION: The proposed model improves the only previously published work on gonio-photographs segmentation and may be a valid support for the automatic processing of these images to evaluate local tissue morphology. Uncertainty estimation is expected to facilitate acceptance of this system in clinical settings.

On Clinical Agreement on the Visibility and Extent of Anatomical Layers in Digital Gonio Photographs.

Purpose: To quantitatively evaluate the inter-annotator variability of clinicians tracing the contours of anatomical layers of the iridocorneal angle on digital gonio photographs, thus providing a baseline for the validation of automated analysis algorithms. Methods: Using a software annotation tool on a common set of 20 images, five experienced ophthalmologists highlighted the contours of five anatomical layers of interest: iris root (IR), ciliary body band (CBB), scleral spur (SS), trabecular meshwork (TM), and cornea (C). Inter-annotator variability was assessed by (1) comparing the number of times ophthalmologists delineated each layer in the dataset; (2) quantifying how the consensus area for each layer (i.e., the intersection area of observers' delineations) varied with the consensus threshold; and (3) calculating agreement among annotators using average per-layer precision, sensitivity, and Dice score. Results: The SS showed the largest difference in annotation frequency (31%) and the minimum overall agreement in terms of consensus size (∼28% of the labeled pixels). The average annotator's per-layer statistics showed consistent patterns, with lower agreement on the CBB and SS (average Dice score ranges of 0.61-0.7 and 0.73-0.78, respectively) and better agreement on the IR, TM, and C (average Dice score ranges of 0.97-0.98, 0.84-0.9, and 0.93-0.96, respectively). Conclusions: There was considerable inter-annotator variation in identifying contours of some anatomical layers in digital gonio photographs. Our pilot indicates that agreement was best on IR, TM, and C but poorer for CBB and SS. Translational Relevance: This study provides a comprehensive description of inter-annotator agreement on digital gonio photographs segmentation as a baseline for validating deep learning models for automated gonioscopy.

A lattice estimation approach for the automatic evaluation of corneal endothelium density.

The analysis of microscopy images of corneal endothelium is routinely carried out at eye banks to assess cell density, one of the main indicators of cornea health state and quality. We propose here a new method to derive endothelium cell density that, at variance with most of the available techniques, does not require the identification of cell contours. It exploits the feature that endothelium cells are approximately laid out as a regular tessellation of hexagonal shapes. This technique estimates the inverse transpose of a matrix generating this cellular lattice, from which the density is easily obtained. The algorithm has been implemented in a Matlab prototype and tested on a set of 21 corneal endothelium images. The cell densities obtained matched quite well with the ones manually estimated by eye-bank experts: the percent difference between them was on average -0.1% (6.5% for absolute values). Albeit the performances of this new algorithm on the images of our test set are definitely good, a careful evaluation on a much larger data set is needed before any clinical application of the proposed technique could be envisaged. The collection of an adequate number of endothelium images and of their manual densities is currently in progress.