Robust filtering of thin-slice reconstructions improves lacunar stroke detection in CT perfusion imaging.

BACKGROUND

The detection of lacunar strokes in CT perfusion (CTP) imaging may be improved by reducing the reconstructed CTP slice thickness. However, this increases noise, necessitating the use of robust noise filters.

PURPOSE: To Investigate the impact of slice thickness and three state-of-the-art noise filters on the detection of lacunar strokes of varying sizes.

METHODS

Artificial spherical lacunar infarcts (diameter: 5, 7, and 10 mm) were added to 40 thin slice CTP acquisitions. The performance of time-intensity profile guided filtering, temporal-average guided filtering, and a U-Net based filter were compared based on infarct detectability in cerebral blood volume maps of thick- (4.9 mm) and thin-slice (0.7 mm) reconstructions. One observer evaluated the maps for each combination of filter and slice thickness. A second observer repeated the evaluation for the two best performing combinations. F1-score, contrast-to-noise ratio, and observer confidence were calculated for every combination of filter, slice thickness, and observer.

RESULTS

The highest F1-scores were achieved with the temporal-average guided filter (thin slices: 0.70, thick slices: 0.49). The other filters did not exceed an F1-score of 0.36 on either thin- or thick slices. Thin-slice temporal-average guided filtering also achieved the highest contrast-to-noise ratio and observer confidence, which resulted in a sensitivity of 10% for 5 mm, 73% for 7 mm, and 75% for 10 mm diameter lacunar infarcts.

CONCLUSIONS

This study demonstrates that robust noise filtering, particularly temporal-average guided filtering, enables improved detection of lacunar strokes (≥7 mm) in CTP imaging using thin-slice reconstructions.