Role of artificial intelligence in MS clinical practice

Abstract

Machine learning (ML) and its subset, deep learning (DL), are branches of artificial intelligence (AI) showing promising findings in the medical field, especially when applied to imaging data. Given the substantial role of MRI in the diagnosis and management of patients with multiple sclerosis (MS), this disease is an ideal candidate for the application of AI techniques. In this narrative review, we are going to discuss the potential applications of AI for MS clinical practice, together with their limitations. Among their several advantages, ML algorithms are able to automate repetitive tasks, to analyze more data in less time and to achieve higher accuracy and reproducibility than the human counterpart. To date, these algorithms have been applied to MS diagnosis, prognosis, disease and treatment monitoring. Other fields of application have been improvement of MRI protocols as well as automated lesion and tissue segmentation. However, several challenges remain, including a better understanding of the information selected by AI algorithms, appropriate multicenter and longitudinal validations of results and practical aspects regarding hardware and software integration. Finally, one cannot overemphasize the paramount importance of human supervision, in order to optimize the use and take full advantage of the potential of AI approaches.

Keywords: Artificial intelligence; Deep learning; MRI; Machine learning; Multiple sclerosis; Neural networks.

Fig. 1

A DL approach for the classification of CNS diseases from MRI scans (adapted from (Rocca et al., 2020) (permission pending)). First, the CNN algorithm was trained using 3D T1- and T2-weighted images from a training dataset. The algorithm was based on 4 blocks, created for each of the image contrasts and finally concatenated. A final model (f) was then generated after feature extraction, concluding Step 1 (algorithm training). Then, new unseen images from the test dataset were processed using the same architecture and parameters as the trained CNN, and the algorithm (f) was applied. A probability of belonging to each disease class was obtained and the final classification of the patient was the class with the highest probability, concluding Step 2 (algorithm application). Finally, algorithm performance was compared against two expert neuroradiologists blinded to clinical information. The images on the right show two examples of correctly diagnosed images by the CNN but not by expert neuroradiologists. Axial T2-weighted (first row) and T1-weighted (second row) slices are included for each patient. A) is a patient showing T2-hyperintense periventricular lesions (stars), with a diagnosis of NMOSD classified as NMOSD by the algorithm and as MS by neuroradiologists. B) is a patient with migraine, having multiple subcortical T2-hyperintense lesions (stars), correctly classified as migraine by the algotithm but as CNS vasculitis by neuroradiologists. Abbreviations: CNN, convoluted neural networks; MS, multiple sclerosis; NMOSD, neuromyelitis optica spectrum disorders.

Fig. 2

Clinical, MRI and cognitive phenotypes. A. Despite no MRI biomarkers have been validated for separating RRMS from PMS patients in the clinical setting, Fig. 2A illustrates the promising approach of measuring upper cervical cord GM cross-sectional area through manual segmentation on a high-resolution phase sensitive inversion recovery (PSIR) sequence at C2-C3 intervertebral level. An example from a patient with RRMS (left) and a patient with PMS (right) is reported, with blue and red lines indicating spinal cord and GM segmentations, respectively. The latter is used for calculating upper cervical cord GM cross-sectional area. B. (adapted from Eshaghi et al. (Eshaghi et al., 2021)) illustrates the evolution of MRI abnormalities in each of the three identified MRI-based subtypes. The colour shade ranges from blue to pink, with increasing degree probability of abnormality. The cortex-led subtype (left) showed cortical atrophy in the occipital, parietal and frontal cortex in the early stages of the sequences, and a reduction in T1/T2 ratio in the normal-appearing WM in the later stages. The normal-appearing WM-led subtype (middle) showed a reduction in T1/ T2 ratio of the cingulate bundle and corpus callosum in the earlier stages of the sequence, and deep grey matter and temporal grey matter atrophy in the later stages. The lesion-led subtype (right) shows early and extensive accumulation of lesions in the earlier stages of the sequence, and a reduction in the T1/T2 ratio in the normal-appearing WM in the later stages. C. illustrates the study by De Meo et al. (De Meo et al., 2021) identifying cognitive phenotypes. A cohort of 1212 MS patients and 196 HC underwent a standardized neuropsychological battery. Age-, sex- and education-corrected z-scores were derived for each neuropsychological test and a ML approach (latent profile analysis) was used to identify five cognitive phenotypes: preserved cognition, mild verbal memory/semantic fluency, mild multi-domain, severe attention/executive, and severe multi-domain involvement. Abbreviations: RRMS, relapsing-remitting multiple sclerosis; PMS, progressive multiple sclerosis; WM, white matter.

Fig. 3

CNN algorithms for automated MS lesion segmentation. In the upper part of the figure (adapted with permission from Valverde et al. (Valverde et al., 2017)), the algorithm proposed by Valverde et al. based on 3D FLAIR and T1-weighted images was compared with older DL algorithms (showing superior performance), holding manual segmentation as the gold standard. The figure depicts a FLAIR (A) and T1-weighted (B) slice, and WM lesion segmentation mask performed manually (C), by older algorithms (D-E) and by the proposed algorithm (F). On all images, true positives are denoted in green, false positives in red and false negatives with a blue square. Likewise, in the lower part of the figure (adapted with permission from Aslani et al. (Aslani et al., 2019)), the algorithm proposed by Aslani et al. based on 3D FLAIR and T1-weighted images was compared with older DL algorithms (showing superior performance), holding manual segmentation as the gold standard. Each algorithm is illustrated by one column of images: from up to below, axial, coronal and sagittal FLAIR slices, together with WM lesion masks, and 3D lesion masks. On all images, true positives, false negatives, and false positives are colored in red, green and blue, respectively.