Reported Case: DLR-Supported MRI Throughput After One Scanner Removed

After scanner consolidation reduced the University of Oulu’s MRI fleet by one unit, the authors report that applying deep learning reconstruction (DLR) helped the department largely preserve overall throughput. In the post-implementation period described, average hourly throughput across the entire fleet in 2025 declined by 6.4% despite operating with fewer scanners. The report frames this relatively limited fleet-level change alongside faster scan delivery at the suite level, rather than as a single isolated metric. It links the observed throughput preservation to DLR paired with protocol optimization, while noting that image-quality caveats emerged in specific sequence and exam contexts.

The analysis is described as a retrospective operational comparison using MRI scanner log data alongside image assessments performed as part of routine quality assurance. The authors compared 10 months in 2023 before DLR implementation with the same 10 months in 2025 after DLR, during a period when the department was working with one fewer MRI machine. They identify the DLR software as coming from Siemens Healthineers and place the work in the context of a late-2024 departmental consolidation and relocation that reduced capacity and required adaptation to a smaller fleet. For the post-implementation interval, the authors add that deployment and protocol optimization were still underway, positioning the comparisons within a real-world, partial-implementation setting.

Within that pre/post framework, the authors report time changes at two levels: sequence duration and total exam time. For sequence duration, they describe reductions of 5–11 minutes, corresponding to 11.5%–27.2% shorter sequences on optimized scanners. For total exam time, they report reductions ranging from 5 minutes to nearly 11 minutes, corresponding to 9.5%–21.2% shorter exams. The authors attribute these time differences to DLR deployment alongside ongoing protocol optimization and connect the reductions to improved productivity per scanner in the post-implementation period. They present the time measures as the central operational evidence for how the department sustained output while scanning with a smaller fleet.

Alongside productivity measures, the authors describe image-quality observations that they say may limit DLR applicability due to “unpredictable performance.” They report this concern as particularly relevant in neuroimaging, where artifacts were found in T2-weighted sequences, and they also note reduced quality in some contrast-enhanced studies. The authors emphasize the role of rigorous quality assurance when implementing deep learning in clinical practice. They also suggest that future research on radiological technologies incorporate financial and productivity indicators alongside diagnostic value to give decision-makers a more comprehensive view of overall impact. Overall, the write-up pairs reported time-based productivity gains with sequence- and study-type-specific quality caveats and a broader measurement agenda proposed by the authors.

Key Takeaways:

• The authors describe a retrospective comparison of 10 months in 2023 versus the same 10 months in 2025 after a consolidation left the department operating with one fewer scanner, with throughput reported as largely maintained at the fleet level.
• In the pre/post analysis, reported operational changes included shorter sequence durations and shorter total exam times in the post-DLR period.
• The authors report “unpredictable performance,” including artifacts in T2-weighted sequences and reduced quality in some contrast-enhanced studies, and they suggest that future work combine diagnostic evaluation with productivity and financial indicators.