SLO-MSNet: Discrimination of Multiple Sclerosis using Scanning Laser Ophthalmoscopy Images with Autoencoder-Based Feature Extraction

Abstract

AbstractBackgroundOptical coherence tomography (OCT) studies have revealed that compared to healthy control (HC) individuals, retinal nerve fiber, ganglionic cell, and inner plexiform layers become thinner in multiple sclerosis (MS) patients. To date, a number of machine learning (ML) studies have utilized Optical coherence tomography (OCT) data for classifying MS, leading to encouraging results. Scanning laser ophthalmoscopy (SLO) uses laser light to capture high-resolution fundus images, often performed in conjunction with OCT to lock B-scans at a fixed position, removing the effects of eye motion on image quality and allowing for evaluating the disease progression at follow-up examinations. To our knowledge, no ML work has taken advantage of SLO images for automated diagnosis of MS.MethodsIn this study, SLO images were utilized for the first time with the purpose of fully automated classification of MS and healthy control (HC) cases. First, a subject-wise k-fold cross-validation data splitting approach was followed to minimize the risk of model overestimation due to data leakage between train and validation datasets. Subsequently, we used several state-of-the-art convolutional neural networks (CNNs), including VGG-16, VGG-19, ResNet-50, and InceptionV3, as well as a custom CNN architecture trained from scratch. In the next step, we designed a convolutional autoencoder (CAE) to extract semantic features from the images which are then given as the input to four conventional ML classifiers, including support vector machine (SVM), k-nearest neighbor (K-NN), random forest (RF), and multi-layer perceptron (MLP).ResultsThe custom CNN model outperformed state-of-the-art models with an accuracy (ACC) of 85%, sensitivity (SE) of 85%, specificity (SP) of 87%, and AUROC of 93%; however, utilizing a combination of the CAE and MPL yields even superior results achieving an ACC of 88%, SE of 86%, SP of 91%, and AUROC of 94%, while maintaining high per-class accuracies. The best performing model was also found to be generalizable to an external dataset from an independent source, achieving an ACC of 83%, SE of 87%, and SP of 79%.ConclusionFor the first time, we utilized SLO images to differentiate between MS and HC eyes, with promising results achieved using combination of designed CAE and MLP which we named SLO-MSNet. Should the results of the SLO-MSNet be validated in future works with larger and more diverse datasets, SLO-based diagnosis of MS can be reliably integrated into routine clinical practice.