The AI-driven model used in this study was designed to initially analyze three key MRI sequences—T2-FLAIR, diffusion-weighted imaging (DWI), and susceptibility-weighted imaging (SWI) or T2* gradient-recalled echo (GRE). The purpose of this initial analysis was to determine whether additional imaging would be necessary. If no tumor or hemorrhage was identified in this first stage, a secondary model was applied, incorporating a fourth sequence, T2-weighted imaging, to further refine the protocol recommendation (Fig. 1).  

This study analyzed a total of 1,332 cases, which included 362 cases of infarcts, 200 cases of tumors, 192 cases of hemorrhages, and 638 cases categorized as normal or other. The data were collected from multiple healthcare institutions across four countries: Denmark (n=485), India (n=432), Brazil (n=255), and the United States (n=160). The imaging scans utilized in the study were obtained from MRI machines manufactured by three major vendors: Siemens (n=699), GE (n=480), and Philips (n=153) (Fig. 2).  

The overarching goal of this approach was to retrospectively simulate a more efficient MRI protocol selection process, allowing for adjustments while the patient was still in the scanner. By dynamically tailoring imaging protocols, the study aimed to reduce unnecessary scans, optimize workflow efficiency, and streamline the neuroimaging process without compromising diagnostic accuracy.