Utilization of artificial intelligence to diagnose stroke in magnetic resonance imaging

Stroke is a common cause of disability and death worldwide. Fast diagnosis of the etiological cause of stroke is crucial for the patient’s prognosis. The ischemic and hemorrhagic causes cannot be distinguished without a brain scan, and magnetic resonance imaging (MRI) is becoming increasingly popular as the firstchoice imaging for patients with acute stroke. However, increasing radiological workloads risk causing delayed diagnosis or either missed or erroneous diagnosis. 

Deep learning-based Artificial intelligence (AI) algorithms are rapidly advancing, and commercial systems promise to help maintain diagnostic speed and correctly diagnose patients. Commercial algorithms must demonstrate their ability before receiving a European Conformity (CE) mark. This usually consists of an internal diagnostic test. Nevertheless, the external reproducibility of the internal test results has yet to be investigated.  

This PhD thesis aimed to assess the feasibility of using artificial intelligence for the detection and mismatch classification of lesions on MRI scans from patients suspected of stroke. 

The Systematic Review evaluated all available scientific research, and we concluded that the current AI technology could confidently diagnose ischemia in MRI with a sensitivity and specificity of 93% and 93%, respectively. Furthermore, we determined that limited evidence existed for diagnosing hemorrhage. Lastly, the study highlighted that only one CE-marked AI solution had a scientific publication that examined its detection ability.

The Detection Study investigated a commercially available AI algorithm solution for detecting MRI lesions compatible with stroke in a comprehensive stroke center treating all stroke types. The study examined the AI's ability to categorize scans based on lesion presence and its ability to detect single ischemic and hemorrhagic lesions. The study found a significant difference in sensitivity between the categorization of scans and the detection of single lesions. This study concluded that ischemic and hemorrhagic lesions can be identified but are not an optimal solution for use in a comprehensive stroke center.

The Mismatch Study examined an improved stroke-specific commercial AI algorithm with integrated mismatch evaluation of diffusion weighted imaged (DWI) and fluid attenuated inversion recovery (FLAIR) sequences that could be utilized for decision aid in the acute ischemic stroke type. Five expert evaluators were used as comparators to establish a ground truth and for interrater agreement analyses. The study found that scans with ischemia and scans with hemorrhage were able to be detected. The mismatch classification of ischemia was significantly worse than explainable by any interrater variance, although an AI low mismatch classification could be used for ruling out high mismatch conditions.

The combined studies in this thesis conclude that AI can be used to assess stroke-suspected patients, and prospective trials should be conducted to investigate.