E2378. Impact of Deep Learning Image Reconstruction Methods on MRI Throughput in an Outpatient Setting
Authors
Mark Finkelstein;
Mount Sinai Hospital
Anthony Yang;
Mount Sinai Hospital
Clara Koo;
Mount Sinai Hospital
Amish Doshi;
Mount Sinai Hospital
Objective:
Although artificial intelligence (AI) based methods have been shown to improve MRI image quality while reducing scan time, it has yet to be shown whether they can generalize beyond an individual scanner or research setting and into realistic clinical practice (1). In this study, we assess the ability of two such implementations to increase workflow efficiency, productivity, and profitability by using DL-enhanced fast protocols on real-time MR examinations performed at multiple outpatient facilities in a large multicenter institution.
Materials and Methods:
The institutional Radiological Information Service (RIS) was queried for all MRI scans performed at three outpatient sites in January 1, 2019 to July 31, 2019 and January 1, 2022 to July 31, 2022, with room time and examination type extracted. This data was linked with scan time data obtained from the picture archiving and communication system (PACS). Scanners were categorized based on utilized AI method (SubtleMR, AIR Recon DL, and none). Comparison was performed between the scan time and room time for different exam types, categorized by exam codes, between scanners grouped on utilized AI method. Additionally, no scanners utilized an AI method in 2019, thus for scanners utilizing an AI method, this year was used to evaluate the direct intervention of an AI method on these scanners. Scans that involved multiple exam codes were excluded. Exam codes were linked to expected RVUs to calculate the expected per minute RVU savings that may be expected to be obtained from the utilization of an AI method.
Results:
A total of 40,580 scans across 9 MRI scanners were analyzed. Three scanners utilized SubtleMR, two utilized AIR Recon DL, and four did not have AI implemented. Substantial reductions in scan and room time were seen for certain study types, for example for MRI examinations of the knee without contrast with SubtleMR in comparison to scanners with no AI method implemented, a 65% reduction (p < 0.001) in scan time, corresponding to a mean increase in relative value units (RVU) of 31.2/hour, and 32% reduction (p < 0.001) in room time, corresponding to a mean increase in RVU of 5.4/hour. There were also studies that exhibited a clear decrease in scan time, without a significant corresponding decrease in room time, for example MRA Head without contrast, for which scan time was seen to decrease by 47% (p < 0.001), however room time was only found to decrease by 3% (p = 0.54).
Conclusion:
Utilization of AI methods has the potential to add value to radiology practice groups by facilitating increased throughput. Our study exhibited clear decreases in scan times as well as the more relevant statistic of room time for certain study types. However, these potential time savings are not seen for all study types. Furthermore, though there may be an apparent decrease in scan time, this may not correlate to a decrease in room time, likely indicate that factors other than the scan time dictate study length. This would indicate that potential implementations must consider practice mix in consideration of different AI methods.
