Diffusion motion correction boosts portable brain CT image quality

A diffusion model-based algorithm boosts image quality and diagnostic confidence for portable brain CT images without compromising lesion detection, according to a study published January 6 in Academic Radiology.

The finding underscores the algorithm’s potential for clinical use, especially for severely ill patients who can benefit from portable CT, either in the field or at bedside, noted a team led by Zennong Chen, PhD, of Massachusetts General Hospital and Harvard Medical School, both in Boston.

“Patient motion is a major source of artifacts in portable brain CT … [as] critically ill patients often lack the ability to voluntarily stabilize their head, and existing immobilization devices may not be appropriate for such patients,” the group explained.

To address the problem, some have proposed a diffusion model-based algorithm for portable CT, according to the investigators, who wrote that “the diffusion model is conditioned on motion-corrupted images and trained to estimate the score function of the corresponding motion-free images.” The team sought to evaluate the performance of this diffusion model-based motion correction algorithm for portable brain CT via a study that included 67 portable brain CT scans that also had corresponding fixed CT scans taken within two days as reference.

The investigators applied a pretrained diffusion model to correct motion artifacts in the portable scans; each of the 67 cases yielded three volumes: original (motion group), corrected (corrected group), and fixed (reference group). Three readers — one neuroradiologist, one neuroradiology fellow, and one radiology resident — reviewed the images in random order and scored eight lesion types and four image quality metrics on a 5-point Likert scale.

The team found the following:

Image quality scores across four portable brain CT exam metrics on a 5-point Likert scale
Metric	Original exam group (i.e., motion)	Corrected exam group	Reference exam group	p-value
Overall quality	2.28	2.61	3.44	< 0.001
Sharpness	2.48	2.51	3.34	0.34
Motion artifacts	2.43	3.22	3.91	< 0.001
Diagnostic confidence	2.52	2.86	3.69	< 0.001

“Corrected images significantly outperformed motion images in all image quality metrics … except for sharpness,” they wrote.

The authors also reported that lesion detectability was comparable before and after correction, with no significant differences in agreement rates or area under the receiver operating curve (AUC) across all lesion types, and that compared with the reference group, agreement rates ranged from 0.866 to 0.985 and AUCs from 0.788 to 0.964.

Visual examples of image quality improvements. (A) and (B) show corrections of motion artifacts within the brain tissue, where (A) corrects the star-like artifacts originating from the skull and (B) corrects the severe streaking artifacts across the brain. (C) shows the removal of a “double skull” artifact caused by substantial head motion. Image display window is [0,80]HU for (A) and (B) and [−500, 1500]HU for (C). Images and caption courtesy of Academic Radiology.

The study results suggest that “this [motion correction] approach holds clinical potential by providing radiologists with cleaner, more interpretable images and increasing their confidence in making accurate diagnostic assessments,” the authors concluded.

Access the full study here.