Lesion number and burden can predict the long-term outcome of multiple sclerosis, while the localization of the lesions is also a good predictive marker of disease progression. These biomarkers are used in studies and in clinical practice, but the reproducibility of lesion count is not well-known.
In total, five raters evaluated T2 hyperintense lesions in 140 patients with multiple sclerosis in six localizations: periventricular, juxtacortical, deep white matter, infratentorial, spinal cord, and optic nerve. Black holes on T1-weighted images and brain atrophy were subjectively measured on a binary scale. Reproducibility was measured using the intraclass correlation coefficient (ICC). ICCs were also calculated for the four most accurate raters to see how one outlier can influence the results.
Overall, moderate reproducibility (ICC 0.5–0.75) was shown, which did not improve considerably when the most divergent rater was excluded. The areas that produced the worst results were the optic nerve region (ICC: 0.118) and atrophy judgment (ICC: 0.364). Comparing high- and low-lesion burdens in each region revealed that the ICC is higher when the lesion count is in the mid-range. In the periventricular and deep white matter area, where lesions are common, higher ICC was found in patients who had a lower lesion count. On the other hand, juxtacortical lesions and black holes that are less common showed higher ICC when the subjects had more lesions. This difference was significant in the juxtacortical region when the most accurate raters compared patients with low (ICC: 0.406 CI: 0.273–0.546) and high (0.702 CI: 0.603–0.785) lesion loads.
Lesion classification showed high variability by location and overall moderate reproducibility. The excellent range was not achieved, owing to the fact that some areas showed poor performance. Hence, putting effort toward the development of artificial intelligence for the evaluation of lesion burden should be considered.