Multiple sclerosis (MS) is a chronic autoimmune inflammatory and demyelinating disease of the central nervous system. While its etiology is still unknown (Polman et al., 2011), MS is the first cause of non-traumatic neurological disability in young adults, affecting about 2.8 million people worldwide (Goodin, 2014). Often starting with a preliminary clinical isolated syndrome (CIS) involving a large heterogeneity of clinical symptoms such as weak limbs, blurred vision, dizziness, fatigue, or tingling sensations, the disease may evolve along two main clinical courses. In 85% of patients, the disease starts as a relapsing-remitting course (RRMS, noted RR), with the occurrence of relapses. These RRMS patients can evolve over time into a non-systematic secondary-progressive course (SPMS, noted SP). In the 15% remaining patients, the disease evolves as primary-progressive MS (PPMS, noted PP) which corresponds to a steadily worsening of symptoms over time without any relapses (Lublin et al., 2014). The current McDonald diagnostic criteria for MS combine clinical assessment, imaging, and laboratory findings (Thompson et al., 2018). Despite such clinical classification, the status and the evolution of each patient could be very different from one to another, leading more and more to individual therapeutic approaches. Thus, to propose personalized medical care and therapy, the neurologist needs to better predict the disease evolution based on early clinical, biological, and imaging markers available from disease onset.

Magnetic Resonance Imaging (MRI) is the most effective tool for the diagnosis of MS and for monitoring the disease modifying treatment. Conventional MRI provides T1-weighted (T1w), T2-weighted (T2w) and FLAIR images allowing the detection and follow-up of white matter (WM) lesions for clinical care (Mure et al., 2016). These conventional sequences allow the quantification of whole brain, WM or gray matter (GM) atrophy using dedicated software. More advanced MRI sequences such as diffusion-weighted imaging (DWI) and diffusion tensor imaging (DTI) have been developed to provide more sensitive markers of the inflammation processes occurring in WM and leading to T1- and T2-lesions. Several metrics of DTI such as the fractional anisotropy and the mean diffusion enable the detection of micro-architectural alterations in WM lesions as well as in normal-appearing WM (Jutten et al., 2019).

More recently, graph theory methods have been used to model brain network organization (Rubinov and Sporns, 2010; Guo et al., 2017). These graph models consist of nodes, based on the parcellation of brain GM regions, and edges, determined by the underlying links between the network nodes. In brain structural connectivity, these links are defined by the extraction of WM fibers using DTI tractography (Hagmann et al., 2007). Previously, Kocevar et al. (2016) have demonstrated an interest of such approaches for the classification of MS clinical profiles using Machine Learning (ML) methods, while Marzullo et al. (2019) improved the classification performance by a first approach using a Deep Learning (DL) model.

However, DTI data used for structural connectivity modeling require long acquisition time and complex processing techniques, which limits its applicability in clinical practice. Nevertheless, brain connectivity can also be obtained from conventional MRI by measuring morphological metrics of the GM on T1w images (Raamana and Strother, 2018). Indeed, several imaging investigations have shown that GM atrophy is present early in MS (Durand-Dubief et al., 2012; Eshaghi et al., 2018). Narayana et al. (2013) has found significant cortical thinning in RRMS patients compared to healthy subjects. Hence, the GM degeneration used in brain morphological connectivity models could provide a sensitive marker of the disease evolution. In such graphs, nodes represent GM areas obtained from the GM tissue parcellation, while edges represent a degree of (dis-)similarity between nodes features like GM thickness or curvature (MacDonald et al., 2000). Such approach has been recently used in Alzheimer's Disease (AD), showing that GM network measures predicted hippocampal atrophy rates in preclinical AD, in contrast to other AD biomarkers (Dicks et al., 2020). Also, Mahjoub et al. (2018) proposed to use morphological connectivity to discriminate late mild cognitive impairment from AD patients. Several studies of GM morphological network were used in Autism Spectrum Disorder (ASD) patients. Kong et al. (2019) proposed an auto-encoder-based deep neural network to identify ASD patients from typical controls, while Corps and Rekik (2019) used morphological networks to estimate the ASD patients' age and deduce the age-related cortical regions. In MS, Muthuraman et al. (2016) analyzed morphological GM thickness networks to classify CIS and RRMS patients using the Support Vector Machine model, obtaining a good level of accuracy. Meanwhile, several studies used graph metrics of GM networks to characterize MS patients. Hawkins et al. (2020) found reduced global efficiency and a more random network in RRMS subjects with cognitive impairment. Likewise, lower node degree and connectivity density were found by Rimkus et al. (2019) in MS patients with cognitive impairment. Rocca et al. (2021) combined functional connectivity and GM network to predict clinical worsening in MS, confirming that GM atrophy is an important predictor for the conversion from RRMS to SPMS. By using the source-based morphometry approach to decompose the cortical thickness map into different patterns, Steenwijk et al. (2016) have further shown that several anatomical patterns are strongly associated with clinical dysfunction in MS patients. Meanwhile, several studies also addressed the problem of age/gender and cortical thickness correlation, and removed their effects before further analysis. Eshaghi et al. (2016) fitted the linear regression between age and GM measurements and took only the residual part to classify MS cohort from neuromyelitis optical patients. Given the graph nature of brain connectivity, the use of graph neural network (GNN) to process such data is an evitable path. GNN allows us to deal with the heterogeneity of input data by capturing the message passing across nodes (Bronstein et al., 2021). More specifically, graph convolutional network (GCN), a reimplementation of convolution concept on GNN, is now ubiquitous in solving problems on non-euclidean data.

In the meantime, the application of convolutional neural network (CNN) has proven its strong ability in computer vision, especially in the biomedical image processing field. Leclerc et al. (2019) has successfully delineated cardiac structure on ultrasound images through an encoder-decoder-based model. 3D-CNN, a particular type of CNN, has been widely used in medical context since a huge amount of medical images were acquired and reconstructed in 3 dimensions. Various studies have focused on disease detection from anatomical neuroimaging (Wargnier-Dauchelle et al., 2023). Huang et al. (2019) have built a VGG-like CNN to adapt 3D image challenge for the purpose of Alzheimer's Disease (AD) classification using both T1w-MRI and FDG-PET modalities for a better outcome. Folego et al. (2020) have adapted LeNet, VGGNet, GoogLeNet, and ResNet in 3D domain to the aim of AD detection. Flaus et al. (2022) has proposed a 3D sequential ResNet to enhance PET images for better visualization of brain lesions. A transparent CNN framework proposed by Eitel et al. (2019) has revealed the decision process of CNN in the diagnosis of MS and pointed out more disease-relevant features in MR images. Optic nerve lesions, one of the first manifestations of MS, can be detected by the 3D-CNN model designed by Marti-Juan et al. (2022).

In this study, we proposed to use GCN for the classification of MS clinical forms based only on the measurement of GM morphological feature (thickness) obtained from T1w-MRI. The impacts of different methodological parameters such as the spatial resolution of the GM parcellation atlases and the level of different graph thresholds were compared. Finally, in order to demonstrate the interest of GCN for MS clinical forms classification, we compared the GCN with a classic 3D-CNN approach.

Our method was divided into three steps: (i) cortical feature extraction using FreeSurfer (Fischl, 2012); (ii) generation of brain morphological graphs using distance computation and threshold; and (iii) clinical forms classification using GCN.

In this section, we first present the GCN classification tasks and then the results without age and gender normalization to allow the comparison with 3D-CNN classification results. Second, the GCN classification results with age and sex normalization are presented.

Graph Convolutional Network is an innovative approach for the classification of clinical forms in multiple sclerosis. While functional and structural connectivities were previously used and provided good results (Ktena et al., 2018; Marzullo et al., 2019), they were constrained by the small size of the database available in clinical routine. To overcome this limitation, one approach is to develop a morphological connectivity method requiring only anatomical T1w MRI for brain studies. In order to test such a hypothesis, we developed a complete pipeline using morphological connectivity and graph convolutional networks. To our knowledge, this is the first attempt to use this approach for the classification of MS clinical forms. Brain graphs were established based on Desikan-Killiany, Destrieux, and Glasser atlases, for GM parcellation. Rejection rates of 60, 70, and 80% were applied to connectivity graphs to preserve solely main differences across brain regions. Morphological connectivity data were fed into GCN while 3D brain images were loaded in 3D-CNN to compare the two classification approaches.

First, non-normalized GCN was compared to 3D-CNN, which was unable to normalize age or gender based on image data. Generally speaking, GCN has outperformed 3D-CNN on 4 out of 5 predefined tasks when the threshold/atlas pair was carefully chosen. For the task RR vs. PP+SP, the F1-score generated by GCN was slightly weaker than the result of 3D-CNN with a 1.8 percentage point. However, it requires more computation resources to train a simple 3 convolutional layers network. In our case, GCN only took 5 h for network training while achieving a better result than 3D-CNN which took more than a week on the same computer. The proposed pipeline has gained in computation time thanks to its dimension-reduction ability. Instead of working on 256 × 256 × 256 volumetric images, the graph approach allowed us to use the adjacent matrix of size 360 × 360 in the most complex case.

The comparison of the two classification networks has also given us insights into the medical image processing field. In general, clinical image classification tasks can be easily affected by acquisition changes (manufacturers, centers, MR field, etc.). In particular, CNNs are sensitive to intensity changes with the use of convolution layers. To address this problem, CNN classification networks must be trained on a large number of images that represent both the variability of the acquisition process and the diversity of the patients. Since most medical datasets are composed of a small number of patients, CNN doesn't usually generate well due to its data-thirsty characteristic. In contrast, GCN can be trained on brain graph features that are less sensitive to image intensity changes. Indeed, cortical thinning is an important biomarker of the MS neurodegenerative process that is visible in T1w images (Narayana et al., 2013). With a brain graph generated from cortical thickness, these small changes in the brain were well-captured by the proposed GCN pipeline. Our pipeline returns a clearer relation between brain atrophy and clinical forms, compared to the 3D-CNN approach, which could be improved by using Grad-CAM (Selvaraju et al., 2020) or similar methods.

Second, normalized GCN was used to classify MS clinical forms. This is essential for clinical forms classification. Binary and multi-class classifications were performed between the three clinical forms (RR, PP, SP). The result of normalized GCN showed that GCN can return satisfactory results on binary classification between MS clinical courses. More specifically, the automatic separation of inflammatory forms from neurodegenerative forms, RR vs. SP and PP groups, has been carried out. The best F1-score was found when separating RR from PP patients, and a good result was also obtained in the RR/SP classification task. On one hand, RR patients present relapses corresponding to focal inflammatory processes. On the other hand, SP and PP patients share the experience of progressive clinical evolution, associated or not with inflammatory activity, resulting from degenerative phenomena of the gray matter. Thus, by grouping SP and PP patients, an adequate result was found when the finest atlas (Glasser) was applied.

The three-class classification is a difficult multi-class categorization task which is further worsened by the imbalanced data distribution. Nevertheless, a promising result was obtained using the Glasser parcellation atlas with a high rejection rate, indicating the advantage of dimension reduction when facing complex brain data such as our case.

Classification of SP and PP was the hardest binary classification task to be accomplished. this is partially due to the small amount of PP cases. Indeed, SP and PP are two neurodegenerative forms sharing similar pathological processes. Moreover, PP is a starting clinical form that can be divided into subclasses depending on the level of disability. With an EDSS score ranging from 2 to 7.5, our PP population is composed of both early and late stages of the disease. The latter ones are more relevant and probably more similar to SP patients as shown in the disease duration at scan. This large variability of disability scores reflects different progressions of the disease and thus different stages of brain alterations. Thus, the SP and some PP patients may share MRI phenotypes which makes the classification difficult, and perhaps even unnecessary.

Achieving good results, the binary classification of HC vs. MS patients was not our primary goal. In general, MS patients can be easily distinguished from healthy subjects in both clinical and imaging ways. In our experience, an F1-score of 100% was observed in all GCN outputs, meaning that all combinations of atlases and thresholds provided enough information for the classification task. Similar results were obtained in the previous work of Marzullo et al. (2019) on brain structural connectivity. Marzullo et al. (2019) has performed the test of HC vs. CIS+RR (24/253) and the test of HC vs. SP+PP (24/325) and achieved the best result (F-measure = 1), demonstrating an evident difference between HC and MS brain morphological and structural networks, respectively.

To further compare our work with other studies, we analyzed the results obtained from Marzullo et al. (2019) and Barile et al. (2022). Apart from the binary classification of HC vs. MS patients, Marzullo et al. (2019) have also tested the separation between early and progressive forms of MS (CIS+RR vs. SP+PP: 253/325) obtaining the highest F-measure at 0.99. Since CIS subjects are included in the RR group in our study, we can compare the previous result with our classification task of RR vs. SP+PP (299/361), leading to an F1-score of 0.678. This strong difference in performance demonstrates that white matter inflammation introduced significant information that facilitates the classification of clinical forms in MS. In contrast, the work of Barile et al. (2022) was performed on GM morphological connectivity. Three similar tasks were reported: (1) CIS+RR vs. PP; (2) CIS+RR vs. SP; (3) CIS+RR vs. SP+PP. By employing the same pipeline of graph generation and atlas (Glasser) and an ensemble of machine learning methods, they have obtained an F1-score of 0.661 (0.12), 0.654 (0.12), 0.648 (0.11) for the three tasks, respectively. In our study, we obtained better F1 scores of 0.671 (0.117), 0.711 (0.107), 0.678 (0.063) for the same tasks. This gain in performance (higher F1-score and reduced standard deviation) demonstrated the interest of brain graph convolutional networks.

Taxicab distance is an L1-norm metric that is generally preferred over Euclidean distance for high-dimension data analysis (Aggarwal et al., 2001). However, since every dimension (mean, standard deviation, skewness, kurtosis) has the same attribution in the calculation of Taxicab distance, our feature vector of four dimensions could not have the same impact on the final value due to the difference in magnitude. In such cases, Mahalanobis distance can overcome the problem while removing redundant information from correlated variables. Since distance measurement was included as edge weight in the input data of GCN, the choice can surely affect the final result. Thus, it is not surprising to observe a better result with Mahalanobis distance supporting the graph generation.

Finally, this work presents several methodological limitations. First the classification results were biased by the class imbalance of the database and the insufficient number of patients. Since the current database consists of a series of multiple MR scans per patient, it does not cover enough variability of the disease, meaning a lack of global vision of the disease. Hence, even if we carefully stop the network training before overfitting, it is hard to extract sufficient features of each MS clinical course to classify an unseen patient by the proposed network, resulting in bad output in some cases. Nevertheless, our cohort study had no bias related to the protocol acquisition, which is unique, guaranteeing the homogeneity of the data. In contrast, a multi-center study is more variable and therefore requires a precise study and corrections of bias.

Although studies on MS mainly focus on white matter and lesion analysis, morphological change in gray matter is a non-negligible aspect of the disease. A full pipeline was proposed in this study for the classification of MS clinical forms. It starts from automatic GM segmentation and surface parcellation, followed by GM thickness analysis using three different granularity of atlases, two different distance measurements, and two different age-gender normalization methods. Thus, a brain resulted in a morphological connectivity graph accompanied by a feature matrix per graph. Four rejection rates corresponding to noise elimination were applied to the graph. A graph convolutional network was performed on these graphs to exploit the hidden information behind GM morphological features. In parallel, a classic 3D convolutional neural network was applied to the brain MRI directly for comparison. The best results were generated by proportional GCN that trained on Glasser parcellation-based graphs with Mahalanobis distance measurement and 80% rejection rate. In future studies, to fully exploit its capacity for clinical image analysis, our method can be implemented on a larger database to predict patients' disease evolution and obtain the correlation between images' information and patients' disability. However, to work with such a heterogeneous study will require developing more advanced graph networks (i.e., with attention) to limit biases such as gender, age and acquisition systems.

The sources code and graph data supporting the conclusions of this article can be found at: https://gitlab.in2p3.fr/thomas.grenier/msgcn-classification.

The studies involving humans were approved by Local Ethics Committee (CPP Sud-Est IV) and French National Agency for Medicine and Health Products Safety (ANSM). The studies were conducted in accordance with the local legislation and institutional requirements. The participants provided their written informed consent to participate in this study.

EC: Conceptualization, Investigation, Software, Writing—original draft, Data curation, Formal analysis, Methodology, Validation, Visualization, Writing—review & editing. BB: Software, Writing—review & editing, Data curation. FD-D: Funding acquisition, Supervision, Writing—review & editing, Validation, Visualization. TG: Methodology, Supervision, Validation, Writing—review & editing, Conceptualization, Formal analysis, Investigation, Visualization. DS-M: Conceptualization, Funding acquisition, Methodology, Project administration, Supervision, Validation, Writing—review & editing, Resources.