Arteriole and venule segmentation in infra-red scanning laser ophthalmoscope (IRSLO) images: a novel dataset and deep learning model

Abstract

Purpose : Infra-red scanning laser ophthalmoscope (IRSLO) images are frequently captured together with optical coherence tomography scans, however, despite excellent vessel contrast and high resolution (see Figure 1), they are rarely leveraged for retinal vascular analysis. Accordingly, we sought to develop a deep learning model to segment the optic disc, arterioles and venules appearing in IRSLO images, and to contribute a new dataset of images with ground-truth annotations to enable further development of computational techniques.

Methods : We annotated 30 IRSLO images captured with the Heidelberg SPECTRALIS platform from two studies (Future-MS, a multiple sclerosis study, and i-Test, a study examining pregnancy, 15 images each). This dataset included a mixture of left and right eyes and macula- and optic disc-centred images. Each image was annotated by one of three masked graders. A subset of 6 images were annotated by all three graders to calculate inter-grader agreement. We used a UNet structure with a ResNet-101 backbone to perform vessel segmentation. A subset of 6 images were kept separate for testing.

Results : On the test set, our method achieved an area under the receiver-operator characteristic of 0.992 for overall vessel detection, 0.986 for arteries, 0.986 for veins, and 0.998 for optic disc segmentation (Table 1). For Dice scores see Table 1; see Figure 1 for an example segmentation. We also demonstrate excellent pairwise agreement between masked graders (Dice score > 0.93 for all comparisons, with ~0.99 for vessels) (Table 1).

Conclusions : We have developed a deep learning model which automatically segments the arterioles, venules, and optic disc in IRSLO images, showing excellent performance. The output could be used to calculate vessel metrics such as fractal dimension, which in other retinal imaging modalities have demonstrated clinical insight into conditions such as stroke and dementia, major causes of morbidity and mortality. In future work we intend to increase our ground truth sample size, incorporate fovea detection and metric calculation to the model, and make the data and model publicly available.

This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.

 

Segmentation results for a single test set image (Mean Dice similarity coefficient across classes = 0.913).

 

Model performance and inter-grader agreement (Dice similarity coefficient).