Analyzing Retinal Vessel Morphology in MS Using Interpretable AI on Deep Learning-Segmented IR-SLO Images

Abstract

Multiple sclerosis (MS), a chronic disease of the central nervous system, is known to cause structural and vascular changes in the retina. Although optical coherence tomography (OCT) and fundus photography can detect retinal thinning and circulatory abnormalities, these findings are not specific to MS. This study explores the potential of Infrared Scanning-Laser-Ophthalmoscopy (IR-SLO) imaging to uncover vascular morphological features that may serve as MS-specific biomarkers. Using an age-matched, subject-wise stratified k-fold cross-validation approach, a deep learning model originally designed for color fundus images was adapted to segment optic disc, optic cup, and retinal vessels in IR-SLO images, achieving Dice coefficients of 91%, 94.5%, and 97%, respectively. This process included tailored pre- and post-processing steps to optimize segmentation accuracy. Subsequently, clinically relevant features were extracted. Statistical analyses followed by SHapley Additive exPlanations (SHAP) identified vessel fractal dimension, vessel density in zones B and C (circular regions extending 0.5-1 and 0.5-2 optic disc diameters from the optic disc margin, respectively), along with vessel intensity and width, as key differentiators between MS patients and healthy controls. These findings suggest that IR-SLO can non-invasively detect retinal vascular biomarkers that may serve as additional or alternative diagnostic markers for MS diagnosis, complementing current invasive procedures.

Keywords: deep learning; feature extraction; feature importance; machine learning; multiple sclerosis; scanning laser ophthalmoscopy; segmentation.

Figure 1

Overview of the proposed method to analyze the morphological changes in SLO images related to MS.

Figure 2

Visualization results for optic disc and cup segmentation. The columns from left to right display original sample IR-SLO images, pre-processing step for disc segmentation, optic disc candidates, post-processing step for disc segmentation, windows surrounding the segmented optic disc, and finally, the optic disc and cup segmented by green and red circles, respectively. Note that the last column shows the zoomed state of the yellow window on the original SLO images in the first column.

Figure 3

Visualization of the vessel segmentation results for four samples, where the columns from first to last indicate the original SLO images, the segmentation algorithm outcomes, the region growing outcomes, the missing algorithm results, and the zoomed-in results of the missing stage for a small window of the images, respectively.

Figure 4

Overview of the SHAP values for the designed XGBoost model; (a) illustrates the contribution of the features for each observation, while (b) shows how each feature impacts the model output, positively or negatively.

Figure 5

Overview of the important features measured by the SHAP method.

Figure 6

Visualization of FI on training, validation, internal test, and external test datasets using SHAP approach.

Figure 7

Visualization of important features obtained from the entire set of clinical features across training, validation, internal test, and external test datasets, in the absence of the feature selection stage. The identified important features are approximately in the same order as the important feature set obtained in Section 3.4.

Figure 8

The distribution and relationship between the first five important features obtained by SHAP method. Each point represents a sample: red dots correspond to MS patients (label = 1), and blue dots correspond to normal controls (label = 0).

Figure 9

Visualization of five important features in 2-space dimension using t-SNE. Label 0: HC images, label 1: MS images.

Figure 10

Representation of the first five important features on the images with MS (first row) and the HC images (second row).