To compare the detection of new/enlarging T2 lesions between semi-automated computer detection and human detection.

New/enlarging T2 lesion counts derived from human radiologist and semi-automated computer software (NeuroRx, Montreal, Canada) were obtained from 887 brain MRI scans from 244 subjects enrolled in the SPRINT-MS trial. A single, blinded radiologist conducted the human interpretation; semi-automated analysis was performed on registered images, using a Bayesian-driven probabilistic classification, followed by random-forest based classification, followed by human visual confirmation (Elliott et al, IEEE Trans Med Imaging, 2013). Spearman’s correlation and Cohen’s kappa were derived.

There is discordance between semi-automated and radiologist detection of new/enlarging T2 lesions. Semi-automated detection reported both a higher proportion of scans with new/enlarging T2 lesions and a greater number of lesions per scan. This discordance is important for clinicians to recognize when applying clinical trial data to clinical practice, which typically uses radiologist readers without the aid of automated lesion detection.