This is a retrospective analysis of 35 consecutive patients diagnosed with MS who underwent baseline MRI between April 2018 and December 2020 and had at least 3 follow-up (FU) - MRIs available. MRIs included 3D-Flair sequences and 3D-T1- sequences after administration of intravenous gadolinium-based contrast. All patients underwent MRIs using Siemens scanners with 1.5 Tesla (Siemens MAGNETOM Avanto FIT 1.5T, Siemens Healthcare, Erlangen, Germany) and 3 Tesla (Siemens MAGNETOM Skyra FIT 3T, Siemens Healthcare, Erlangen, Germany) field strengths.

The study was approved by the local ethics committee. All study protocols and procedures were conducted in accordance with the Declaration of Helsinki.

The data that support the findings of this study are available from the corresponding author upon reasonable request.

All MRI datasets were exported to a research server (NORA Imaging Platform, Freiburg, Germany [14]).

Manual assessment of MRIs was done by a radiology resident (4 years’ experience) and a neuroradiologist (11 years). Results of the reading of the resident were discussed with a consultant as this is the usual process in clinical routine. The time that the resident radiologist and the neuroradiologist needed for assessment of the MRIs was recorded. The time of discussion between the resident and consultant was added, as this represents the usual clinical routine.

All FLAIR sequences of the baseline MRI and 3 follow-up MRIs were assessed for lesions in the following regions: cortical, juxtacortical, periventricular, infratentorial. Additionally, the total number of lesions was calculated and MRIs were rated for new and progressive lesions.

Post-contrast T1-weighted sequences were evaluated for enhancing lesions.

This was done once by the reader alone and once with decision support provided by the automated tool.

The computational pipeline begins with preprocessing of both T1 and FLAIR images using N4 bias field correction [15]. The FLAIR images are then affine co-registered and transformed using ELASTIX [16].

Automatic lesion segmentation is performed using the multi-dimensional gated recurrent units (MD-GRU) algorithm [17], trained on data from 785 patients from the Swiss Multiple Sclerosis Cohort (SMSC).

To locate lesions in space according to the McDonald diagnostic criteria, we employ FreeSurfer [18] and additional post-processing to segment various neuroanatomical structures. Furthermore, we utilize the Hammers Atlas, which includes 95 regions, and condense the subregions into 23 key areas such as the temporal lobe, cingulum, basal ganglia ect. This atlas, in MNI-space, is inversely transformed to T1 space using Dartel from SPM12 [19].

In follow-up studies, images are affine co-registered to the last study. After subtracting the lesion map of both studies, standalone lesions are identified as new lesions, while lesions attached to old ones are defined as progression.

For MS contrast agent enhancement, the script applies a median filter, generates a mask, calculates thresholds, labels connected components, and enhances potential lesions. Vessel images are identified if their intensity exceeds 600, and any overlapping lesions are excluded.

Finally, T1 and FLAIR images, along with the lesion map, ROIs following McDonald diagnostic criteria, and the Hammers Atlas, are visualized on NORA. Lesions and ROIs can be edited if necessary. Based on overlap with ROIs or the atlas, a live report is generated as a table for each study.

To account for the longitudinal nature of the data, with multiple measurements over time for the same patient, we employed linear model with mixed effects. Factors accounted for in the analysis were the patient and the timepoint, coded as baseline, follow-up 1, follow-up 2 and follow-up 3. All reading times were converted as decimal minutes. Statistical testing was performed via ANOVA. For comparing reading times at baseline between neuroradiologist with and without AI-assistance, as all the samples were independent from each other, a paired t-test was used. 95% Confidence intervals were computed via the standard error. The analysis was implemented in R (version 4.2.1, 2022-06-23).