Abstract
Purpose :
NIDEK GS-1 (NIDEK CO., LTD. Japan) collects color images of the Irido-Corneal Angle (ICA) angle split into 16 sectors, each acquired in 17 foci. The manual segmentation of these images could be time consuming and error-prone because the anatomical structures of the ICA have complex boundaries, adjacent tissues are not clearly separated, and some regions may not be in focus. This work proposes a semi-automatic segmentation algorithm based on image processing and prior knowledge of the ICA geometry with the aim of making the user interaction minimal, fast, and effective.
Methods :
A boundary is a sequence of curved lines between seed points (i.e., points on the border between two structures) so that the path connecting them has a minimum cost. The cost function is based on 3 components: the gradients of the RGB input image (1280x960) and the optimizer of a geometrical approximation of ICA’s anatomical structures. The gradients locate the edges and the geometrical component reinforces the weakest contours. The algorithm’s sensitivity to each cost component is automatically evaluated during the segmentation process. A measure of the edges’ strength in the neighborhood of the drawn boundary permits to balance the gradient and the geometrical components. The resulting contour is smoothed with a moving average filter.
Results :
A qualitative evaluation of the results was assessed on a dataset of 40 GS-1 images. For comparison, images were firstly segmented with a manual tool approximating the regions with polylines based on the user-defined seed points. The same images were then segmented with the proposed algorithm using the same seed points (Fig 1). The outcomes demonstrated that our algorithm permits to trace the tissue boundaries more precisely. Moreover, on average, 19% of seed points, frequently located on strong and complex edges, were redundant (Fig. 2). Finally, the effort for the clinician and both the intra- and inter-user variability were reduced.
Conclusions :
The proposed algorithm provides satisfactory segmentations of gonioscopic images and can be also used to collect training data for automatic segmentation and classification algorithms. In fact, it minimizes imperfections due to computer-human interaction, meaning that differences in delineations from multiple experts will only reflect differences in their clinical opinions. Further evaluation should be performed to define a quantitative measure of the algorithm’s performance.
This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.