Evaluation of machine learning algorithms using OCT and OCT angiography for the diagnosis of multiple sclerosis

Abstract

Aims/Purpose: We present a comprehensive evaluation of machine learning (ML) algorithms using Optical coherence tomography (OCT) and optical coherence tomography angiography (OCT‐A) for the diagnosis of Multiple sclerosis (MS).Methods: Prospective observational study using OCT volumes acquired with Cirrus HD‐OCT 5000 (Carl Zeiss, Meditec, Dublin, CA, EEUU) using the scanning protocols Optic Disc Cube 200 × 200 and Macular Cube 512 × 128, and OCT‐A macular volumes 175 × 350 with FOV 6 × 6 mm and OCT‐A of the Optic Nerve Head (ONH) with FOV 4.5 × 4.5 mm. The total data set consisted of 79 eyes of 40 patients with MS and 54 eyes of 27 control subjects. We evaluated classification algorithms such as Support Vector Machines, k‐Nearest Neighbour, Decision Trees, Random Forest, Extra Trees Classifier and Gaussian NB using data extracted from Ganglion Cell Analysis (8 parameters), Macular Thickness Analysis (7 parameters) and Optic Disc Analysis (5 parameters) from the OCT volumes and 5 parameters from the analysis of the macular OCT‐A volumes and 6 parameters from the analysis of the ONH. We studied all the ML algorithms considering three scenarios: OCT (20 parameters), OCT‐A (11 parameters), and OCT + OCT‐A (31 parameters). The models were trained using a 5‐fold cross‐validation strategy. We calculated the accuracy, F1‐score for each model and the area under the curve (AUC) was used to select the best ML algorithm. We applied SHAP (SHapley Additive exPlanations) to explain the output of the best ML model.Results: For the MS diagnosis, the best results were obtained for OCT with Gaussian NB (accuracy 0.865 ± 0.085, F1‐score 0.864 ± 0.086 and AUC 0.871 ± 0.081), for OCT‐A with Random Forest (accuracy 0.832 ± 0.032, F1‐score 0.832 ± 0.043 and AUC 0.831 ± 0.043) and for OCT + OCT‐A with Gaussian NB (accuracy 0.885 ± 0.028, F1‐score 0.886 ± 0.028 and AUC 0.891 ± 0.034). The input parameters with the highest contribution to the model's predictions were the minimum ganglion cell thickness and the temporal flow index in the ETDRS grid.Conclusions: This study highlights the efficacy of machine learning techniques in utilizing combined parameters from OCT and OCT‐A images of the macula and ONH tests to facilitate early diagnosis of MS. Moreover, it underscores the significance of ophthalmic examination in the comprehensive management of the disease.