Background. Artificial intelligence (AI) in medical imaging is a growing and promising technology that can be applied in stroke diagnosis. The study aims to overview studies that compare diagnostic performance of AI applications in stroke detection and seg- mentation of stroke lesions with and without human clinicians, appraising the models, study design, and metrics used.
Materials and methods. This systematic review was performed using the PubMed search engine including articles published in the time frame of 2015 January 1 to 2021 July 23. A to- tal of 438 studies were found, out of which 60 were chosen for the review.
Results. Only 2 out of 60 (3.3%) studies were prospective. Minimum unique computer tomography (CT) scans included for validation – 10, maximum – 21586, mean – 599, me- dian – 100, standard deviation – ±2801.1. The training set sizes consisted of minimum 28 CT scans, maximum – 24214, mean – 1279, median – 153, standard deviation – ±5006.7. Most popular software used in the studies were Brainomix (n=12, 20% of studies) and RAPID (n=12, 20%), 6 studies (10%) used convolutional neural networks, and 6 studies did not iden- tify the model or name of software used. The average value of the ROC AUC results reported was 0.884 and the average accuracy was 0.857. The average reported sensitivity and specific- ity were 0.746 and 0.862, respectively. 27 out of 60 studies used human operators, with the average number of human operators per study being 3.7±2.9.
Conclusions. AI solutions can be widely applied in computation of infarct volumes. Us- ing AI in stroke diagnosis still requires further research with more prospective studies, more expert human operators, and more focus on evaluating secondary outcomes.
This work is licensed under a Creative Commons Attribution 4.0 International License.