Analyzing morphological alternations of vessels in multiple Sclerosis using SLO images of the eyes

Abstract

AbstractBackgroundOur purpose was to investigate the most relevant and discriminating clinical feature set of Scanning laser ophthalmoscopy (SLO) images, which could differentiate multiple sclerosis (MS) and healthy control (HC) patients.MethodsIn this work, SLO images were used for the first time to measure the most valuable manual and clinical features from some retinal structures, optic disc, cup and blood vessels, for MS and HC classifications. For this, first an age-matching algorithm along with a subject-wise k-fold cross-validation data splitting approach were applied for construction of training, validation and test dataset, minimizing the risk of model overestimation. Then, it was needed to segment the retinal structures from the SLO images, and due to the lack of ground truth for our SLO images, we took advantage of a previously proposed deep learning algorithm for anatomical segmentation using color fundus images. But owing to different imaging modalities of SLO images, we also used two stages of pre-processing and post-processing to obtain accurate results for the segmentation step. Following that, a set of manual and clinical features was measured from the segmented optic disc, cup and vessels to gain a better comprehension of the features playing an important role in classification of MS and HC images. Finally, three simple machine learning models were applied to evaluate the measured features and the most valuable and effective features were computed.ResultsThe measured feature set from the segmented optic disc, cup and blood vessels resulted in a mean accuracy (ACC) of 83%, sensitivity (SE) of 79%, specificity (SP) of 85%, and AUROC of 84%, when testing on validation data by using a XGBoost classifier model. Furthermore, horizontally disc location, fractal dimension and intensity variation of blood vessels were selected as the most important and effective features for MS and HC classification.ConclusionThe location of optic disc, fractal dimension and vessel intensity, the ratio between intensity of vessels to intensity of he whole SLO image, were selected as three most valuable features for MS and HC classification. Regarding the optic disc location, we found out the used SLO images had been captured with two different imaging techniques. So, this feature could not be trusted as the most important feature. Two other features were confirmed by one expert as clinically distinguishing features for MS and HC classification.