A systematic review was undertaken to summarize the data regarding accuracy and effectiveness of artificial intelligence algorithms for identifying MRI manifestations of multiple sclerosis. The review included 39 papers, whose authors put forth a multitude of corresponding algorithms and mathematical models. However, quality assessment of these developments was limited by retrospective testing on repeat data sets. Clinical test results were almost entirely absent, and there were no prospective independent studies of accuracy and applicability. The relatively high values obtained for the main measures (similarity, sensitivity and specificity coefficients, which were 7585%) were offset by the methodological errors when creating the baseline data sets, and lack of validation using independent data. Due to small sample sizes and methodological errors when measuring the result accuracy, most of the studies did not meet the criteria for evidence-based research. Studies with the highest methodological quality had algorithms that achieved a sensitivity of 51.677.0%, with a SrensenDice coefficient of 53.556.0%. These numbers are not high, but they indicate that automatic identification of multiple sclerosis manifestations on magnetic resonance imaging may be achievable. Further development of computer-aided analysis requires the creation of clinical use scenarios and testing methodology, and prospective clinical testing.
The quality of artificial intelligence algorithms for identifying manifestations of multiple sclerosis on magnetic resonance imaging (systematic review)
