A prospective single-center non-inferiority study assessed whether Compressed Sensing (CS)–accelerated 3D-FLAIR provides lesion detection comparable to an optimized GRAPPA-accelerated 3D-FLAIR protocol in clinical MRI for multiple sclerosis. The trial enrolled 68 patients with known or suspected MS. Investigators set the primary endpoint as manual lesion count by two blinded readers and applied a 10% non-inferiority margin. The study team reported that CS met non-inferiority for both readers (reader A p < .0001; reader B p = .005). Manual lesion detection therefore did not fall below the predefined performance threshold when CS was used instead of GRAPPA.
Sequence timing differed substantially between the protocols. CS-accelerated 3D-FLAIR shortened acquisition time by 31.8% relative to the optimized GRAPPA sequence; recorded scan times were 4:00 minutes for CS and 5:52 minutes for GRAPPA. The source text did not provide additional technical details such as acceleration factors, CS reconstruction settings, or specific TR/TE values. Those methodological parameters were not reported in the preprint and thus are not described here.
The study team evaluated interrater reliability and significance of non-inferiority. Interrater agreement for manual lesion counts was excellent for CS (ICC = 0.921; 95% CI, 0.837–0.957) and very good for GRAPPA (ICC = 0.851; 95% CI, 0.769–0.906). Both blinded readers produced lesion counts that satisfied the study’s non-inferiority criterion for CS versus GRAPPA. The authors reported p-values for the non-inferiority tests but did not provide additional metrics such as mean lesion counts per patient by sequence or per-lesion concordance tables in the abstract; those details may appear in the full manuscript or supplementary material.
Readers rated subjective image quality as high for both CS and GRAPPA. Objective image metrics diverged: GRAPPA demonstrated higher signal-to-noise ratio (SNR) and contrast-to-noise ratio (CNR) compared with CS. The preprint records these comparative findings but does not include the exact SNR or CNR values within the abstract text. Clinicians should note the dissociation between objective metrics and reader-perceived diagnostic adequacy—subjective quality remained acceptable for CS despite lower measured SNR/CNR.
Automated analyses produced different results from manual reads. The investigative team observed higher automated lesion counts and greater lesion volumes when using CS-accelerated 3D-FLAIR compared with GRAPPA. The abstract did not supply absolute automated counts or volumetric numbers, nor did it detail the software or algorithm used for automated lesion detection. Therefore, specific thresholds, false-positive rates, or voxel-wise comparisons cannot be extracted from the provided source text.
This single-center prospective study received institutional ethics approval from Carl von Ossietzky University, Oldenburg, Germany (No. 2021-033). The trial is registered with DRKS under identifier DRKS00024403. One author declared consulting and lecturing for Siemens on behalf of the employer; the remaining authors reported no known competing financial interests or personal relationships that could have influenced the work. The dataset is not publicly available due to local and European data-protection regulations.
The preprint reports core outcome data but leaves several operational questions open. The abstract omits detailed sequence parameters (e.g., CS acceleration factor, reconstruction settings) and does not present absolute lesion counts per sequence or the numeric SNR/CNR values; these omissions limit reproducibility and direct metric comparison. Automated lesion-detection methods and software were not specified in the abstract, preventing replication of volumetry findings. The study was single-center and performed at 1.5 T; generalizability to other scanners, field strengths, or broader clinical settings therefore remains unaddressed in the source text. Finally, as a preprint, the study has not undergone peer review, and the authors caution that findings should not yet guide clinical practice.
The data that are reported indicate that Compressed Sensing–accelerated 3D-FLAIR reduced scan time by nearly one third and maintained manual lesion-count performance against an optimized GRAPPA protocol, while producing lower objective SNR/CNR and higher automated lesion metrics. Pending access to full-method details and peer review, these results raise concrete questions about algorithmic thresholds for automated lesion detection, the trade-off between objective noise metrics and reader performance, and the behavior of CS protocols across different MRI platforms and field strengths.