Comparative Assessment of Multiple Model Implementation Systems for Multi-Layer Retinal Segmentation: Single Line, Two-Line, and ROI training.

Abstract

Purpose : To evaluate potential performance improvements from using varying levels of global information in deep learning layer segmentation.

Methods : All models were trained with a dataset of 141,000 annotated OCT B-scans, and trained where full images were compressed into 128x128 patches. First, a lowmag model considered the internal limiting membrane (ILM) segmentation alone. A second model used both the ILM and Brooks Membrane (BM) as segmentation targets (two-line model). This focused the model on global OCT architecture. The third model was trained on ILM to BM region of interest (ROI), which then applied the segmentation to the top of the identified region. These models were evaluated on a high-pathology holdout test set of 451 annotated images, and performance was compared using the average deviation away from manual annotations in pixels.

Results : The single line lowmag model had an average offset of 25.38 pixels, heavily impacted by blank sections where the model was unable to detect any ILM line (Figure 1). The two-line model performed better on average than single line segmentations, with an average offset of 2.18 pixels, potentially due to more robust segmentations in high pathology samples. The two-line model also performed better than the ROI model, which had an average offset of 5.65 pixels. Interestingly, the ROI model was less precise than the two-line model, which suggests that there is something useful about both identifying the specific boundaries in the image relevant to the segmentation goal, as well as information regarding the broader context.

Conclusions : Both global and local information are important for making accurate annotations of biological structures which are impacted by pathology and unusual circumstances. This work examines the effect of different levels of global information on deep learning model segmentation performance.

This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.

 

Comparison between the original (top left), and model segmentation methods. Lowmag (top right, yellow), two-line model (bottom left, red) and the ROI model (bottom right, purple) and their average pixel deviation from ground truth (green). Missing undetected sections are circled in red.

 

Average delta between machine learning and ground truth segmentations of ILM.