During recent years, automated diagnosis of MS has been made possible using machine learning algorithms, with a remarkable overall ACC of 94%.
7 Various input data have been utilized thus far, with the most desirable results achieved using the parameters obtained from MRI (pooled ACC = 96%), OCT (pooled ACC = 93%), CSF/serum (pooled ACC = 93%), and even gait and breathing pattern (pooled ACC = 88%) investigations. The OCT-based studies have mainly applied conventional machine learning classifiers on the thickness values
9,10,12,13,15–17,19 or the extracted features from them,
11,14,18 with only two studies utilizing DL.
14 López-Dorado et al.
14 employed Cohen's
d coefficient technique on the thickness maps of RNFL, ganglion cell layer (GCL
+; equivalent to GCIPL), GCL
++ (equivalent to GCIPL plus RNFL), the total retina, and the choroid from 48 MS patients and 48 HC individuals. The resulting thickness map images were then given to a custom CNN model made up of two successive blocks, each containing one convolutional and one pooling layer, ultimately achieving very encouraging results (SE = 100%, SP = 100%). This finding is akin to our results when a similar training and test data splitting approach (record-wise) was employed. Similar to the study by López-Dorado et al.,
14 three other studies captured OCT data using a swept-source device (SS-OCT),
9,10,12 all leading to ACC levels of more than 90%. Such promising results could partly be attributed to the high-resolution scans generated by the SS-OCT technology. It should be noted that these studies had a limited sample size with insufficient diversity; indeed, three studies used the same dataset.
9,12,14 In the second study that utilized DL for distinguishing between MS and HC, Ortiz et al.
27 first analyzed the AUROC curve of the average thickness of both eyes and the inter-eye thickness differences for each of the nine segmented retinal layers (RNFL, GCL, IPL, INL, outer plexiform layer [OPL], outer nuclear layer [ONL], retinal pigment epithelium [RPE], inner retinal layers [RNFL, GCL, IPL, INL, OPL, ONL], and outer retinal layers [RPE and photoreceptors layer]), and identified the most important features accordingly. The GCL average thickness and IPL inter-eye thickness differences were finally selected to be used for training a CNN from scratch. The input size was 8 × 8 × 2, and the model architecture consisted of two consecutive convolutional layers with 16 and 32 kernels (with a size of 3 × 3), generating a 4 × 4 × 32 feature map, which was given to a FC network at the end; an ACC of 87%, a SE of 82%, and a SP of 92% were finally achieved.
27 The largest study that aimed to classify MS based on OCT data was undertaken by Kenney et al.,
19 who evaluated 1568 MS patients and 552 HC subjects from the United States, Europe, and the Middle East. The dataset included various demographic, visual acuity, and SD-OCT parameters; using classification and regression tree models, the authors identified GCIPL thickness of both eyes on average, inter-eye GCIPL thickness difference, and binocular 2.5% low-contrast letter acuity as the features with the highest discriminant capacity. Kenney et al.
19 applied both logistic regression and SVM algorithms that ultimately were shown to have a similar performance. The use of SVM with a linear kernel achieved an ACC of 88%, a SE of 83%, and a SP of 90%. Overall, although the majority of machine learning research on MS classification has taken advantage of MRI,
8 OCT measurements have also been shown to be invaluable input data. Notably, the diagnostic performance of the models trained with MRI and OCT parameters are not far different, but the OCT technology is much less invasive and costly. In the current study, we utilized IR-SLO images in addition to OCT data, resulting in best ACCs of 100% ± 0.0%, 99.19% ± 0.45%, and 96.85% ± 0.45%, respectively, for record-wise, eye-wise, and subject-wise data-splitting approaches. As mentioned above, the proposed bimodal model is indeed a ResNet-101–based CNN with novel FC architecture fitted to our dataset.