AI-based MRI tools show promise in multiple sclerosis diagnosis

A recent Npj Digital Medicine study assesses the accuracy and effectiveness of artificial intelligence (AI)-based imaging techniques to diagnose multiple sclerosis (MS).

Study: A real-world clinical validation for AI-based MRI monitoring in multiple sclerosis. Image Credit: New Africa / Shutterstock.com

Background

MS is a common neurodegenerative and inflammatory demyelinating condition of the central nervous system (CNS). MS is characterized by focal lesions and diffused neurodegeneration in the spinal cord and brain. Individuals with MS suffer significant cognitive and physical disability, which sometimes causes premature withdrawal from the workforce.

Globally, about 2.8 million people are living with MS. Disease-modifying therapy (DMT) has proved to be highly effective and reduces the risk of disease recurrence.

Inflammatory activity is a major pathological substrate that reduces relapse-associated worsening (RAW). The response of MS patients to DMT is annually assessed through magnetic resonance imaging (MRI).

MRI plays a vital role in assessing neurological diseases that affect a large number of axons and disrupt complex integrated brain networks. Likewise, MRI and other imaging modalities facilitate the diagnosis of MS and the monitoring of this disease and its response to DMT.

The lack of prior or current 3D FLAIR volume in picture archiving and communications systems (PACS) poses a limitation for the accurate detection of small lesions. The volume of new or enlarging lesions influences treatment strategies that are not typically detected in routine clinical radiology practice. In traditional methods, radiologists’ experience is extremely important for analyzing the overall FLAIR lesion burden that reflects MS severity. 

The comparison between severe brain volume loss (BVL) and age-matched healthy controls provides significant prognostic information. The accuracy of this information is dependent on the visual inspection of radiologists.

Changes in brain volume during 12-month intervals between MRI scans are small and might not be determined through visual inspection. The inability to identify short-term changes in brain volume is a significant cause of adverse trajectories linked to MS outcomes and influences clinical decisions to change or escalate DMT.

The development of AI algorithms for medical imaging has enabled automation in clinical detection. AI has also been used for the segmentation of brain structures and analysis of different brain pathologies, including MS lesions.

About the study

The current study assessed the effectiveness of iQ-SolutionsTM, hereafter referred to as iQ-MS, based on a large cohort of MS scans. The assessments of MS scans were independently conducted by expert radiologists in clinical settings.

The researchers hypothesized that AI-based tools can more sensitively and accurately evaluate MRI scan reports of disease activity than conventional methods based on radiology reports.

Brain scans were analyzed by iQ-Solutions™ in Digital Imaging and Communications in Medicine (DICOM) format by a collection of AI algorithms based on deep neural network technology. The AI-based algorithms were designed based on 8,500 brain scans that were expertly annotated by skilled neuroimaging analysts.

A reference cohort was created based on MRI scans of over 3,000 healthy controls and an independent sample of 839 people with MS. Both samples were processed with the same methods.

Study findings

The iQ-Solutions^TM system generates data for cross-sectional and longitudinal whole brain, lesion metrics, and brain substructure relevant to MS. This AI tool enables visualization of many picture archiving and communications systems (PACS) for radiologists to review. Scan images are automatically subjected to quality check for optimal pre-contrast 3D-T1 and 3D FLAIR sequences, containing over 30 slices with a thickness of three millimeters (mm) or more.

Cross-sectional segmentation algorithms were designed based on 3D-UNet, which enabled the extraction of image features, followed by the prediction head. Cross-validation was conducted by comparing case- and voxel-wise DICE scores with ground-truth reports produced by skilled neuroimaging analysts.

The lesion activity of different time points was measured by iQ-Solutions, indicating the development of new and enlarging lesions. Moreover, iQ-MS revealed enlarging lesions as new lesioned voxels that are connected to existing lesions reported in a previous study within its 26-voxel neighborhood.

LG-Net is a lesion-inpainting model for brain and substructure volumetric analyses. This system was applied to 3DT1 images to improve the segmentation bias produced due to the presence of MS lesions.

Notably, iQ-Solutions performs many checks between the two scan timepoints. In the event of an error, longitudinal metrics are reported but are returned to the user with a protocol irregularity warning.

The iQ-MS tool is equipped with the DeepBVC algorithm, which assesses longitudinal whole-brain volume change. An AI-based segmentation model integrated with a Jacobian method enabled the estimation of whole gray matter and thalamus volume change. 

Moreover, iQ-MS offers volumetric data for individual patients as normalized values. This tool provides data on brain volumetrics and MS lesion volumes benchmarked to a hypothetical MS patient of similar age, disability, and disease duration. This enabled a more clinically meaningful and experiential reference.

Conclusions

The results support using iQ-MS to monitor people with MS. Compared to a core MRI analysis lab report and radiology reports, the current AI tool offers a better clinical assessment.

The study findings highlight that using iQ-MS could improve clinical imaging, disease-specific research, and the management of individual MS patients in real-time.