AUTHORS: La Rosa F, Lu PJ, Rahmanzadeh R, Cagol A, Barakovic M, Thiran JP, Granziera C, Bach Cuadra M

ECTRIMS 2021 - Multiple Sclerosis Journal, 27(2_SUPPL): 457-458, October 2021


Introduction: Cortical lesions (CL) represent a substantial part of the total lesion burden in multiple sclerosis (MS) patients and are associated with disability progression and cognitive deficits (Treaba, et al., Radiology, 2019). Their visual detection on magnetic resonance imaging (MRI), however, is challenging and requires specialized sequences, such as the magnetization prepared 2 rapid acquisition gradient echoes (MP2RAGE) or the double inversion recovery (DIR).

Objectives: Based on our previous work on automated cortical and white matter lesions segmentation with multi-contrast MRI (La Rosa, et al., Neuroimage Clinical, 2020), we developed and assessed a unimodal architecture (named CLaiMS-Net) based on only the MP2RAGE at 3T to automatically segment CL.

Methods: We considered 80 MS patients (51 female, 29 male) who underwent MRI at 3T (in two Siemens scanners) with MP2RAGE and on which an expert labeled at least one CL. As pre-processing, all images were skull-stripped using HD-BET. CLaiMS-Net was trained and evaluated in a 5-fold cross validation with the MP2RAGE uniform (UNI) image as only input and the CL annotations as ground truth. The white matter lesions (manually labeled as well by an expert) were not considered either in the loss function or as false positives. Additionally, as post-processing step, we discarded all segmented lesions which had a median T1 value of less than 2200 in the MP2RAGE T1 map (the normal T1 of gray matter is in range 1100-1300 ms) to reduce the false positives.

Results: Averaging over the testing folds, CLaiMS-Net achieved a lesion-wise true positive rate (LTPR) of 71%, a lesion-wise false positive rate (LFPR) of 27%, and a Dice coefficient (DSC) of 41%. These values are comparable to the ones of our previous multi-contrast framework, using both MP2RAGE and the fluid attenuated inversion recovery sequence (LTPR of 75% and LFPR of 29%, computed on 90 subjects). A direct comparison between the two, however, is not possible as different subjects were considered in the two studies.

Conclusions: For the first time, we show that a single 3T MRI acquisition (MP2RAGE) is sufficient to obtain an accurate automated CL segmentation with metrics similar to the ones of a state-of-the-art multi-contrast MRI approach.