Predicting new T2 lesions using a machine learning algorithm

A machine learning algorithm was able to detect abnormalities within normal-appearing white matter (NAWM) before any lesion could be detected. Using cross-sectional T1- and T2-weighted non-contrast brain MRI data from NAWM, this machine learning method could predict new T2 lesions up to 48 weeks prior to actual emergence.

Conventional MRI is not sufficiently sensitive to enable early diagnosis, nor is its specificity sufficient to predict disease severity. Machine learning analyses of brain scan data may help to fill this gap. Mr Bastien Caba (Biogen Digital Health, MA, USA) and colleagues analysed brain T1- and T2-weighted MRI scans from the pivotal, phase 3 ADVANCE trial (NCT00906399), which included 1,512 patients with relapsing-remitting multiple sclerosis (RRMS), to validate the algorithm [1]. They then tested this algorithm utilising MRIs of 886 patients with secondary progressive MS (SPMS) who participated in the ASCEND trial (NCT01416181).

Cubic patches with a 15 mm edge were sampled from NAWM of baseline scans. Patches co-locating with a future lesion at 48 weeks post-baseline were labelled positive; patches not associated with a future lesion in spatially matched white matter were negative. Texture-based radiomic features were extracted from the core and periphery of each patch, yielding 372 features per patch.

Of 40 selected features, 22 were core-based and 18 periphery-based; 18 were T1-based, 22 were T2-based. Applied on the ADVANCE validation set, the machine learning algorithm reached 66.4% balanced accuracy, 66.5% precision, 66.0% sensitivity, 66.8% specificity, and an area under the curve (AUC) of 72.6%. In the ASCEND cohort these percentages were 64.6%, 63.7%, 68.0%, 61.2%, and 71.4%, respectively.

“These results further inform our understanding of the nature of lesion formation in multiple sclerosis, which seemingly arises from areas of normal-appearing white matter that are in fact abnormal,” concluded Mr Caba.

1. Caba B, et al. Machine learning-based prediction of new multiple sclerosis lesion formation using radiomic features from pre-lesion normal appearing white matter. S26.009, AAN 2022, 02–07 April, Seattle, USA.

Copyright ©2022 Medicom Medical Publishers



Posted on 2 June 2022