Diffusion Model-Based Motion Correction in Portable Computed Tomography for Brain: A Human Observer Study

Abstract

Rationale and objectives: To evaluate the clinical performance of a diffusion model-based motion correction algorithm for portable brain CT.

Materials and methods: We retrospectively collected 67 portable brain CT scans with corresponding fixed CT scans acquired within ±2 days as reference. A pre-trained diffusion model was applied to correct motion artifacts in the portable scans. Each case yielded three volumes as follows: original (motion group), corrected (corrected group), and fixed (reference group). Images were reviewed in randomized order by three professional readers (one neuroradiologist, one neuroradiology fellow, and one radiology resident), with at least two weeks between sessions to reduce recall bias. Eight lesion types and four image quality metrics were scored using a 5-point Likert scale. ACR phantom testing was performed to assess compliance with diagnostic image quality standards.

Results: Corrected images significantly outperformed motion images in all image quality metrics (improvement: 0.33-0.79, p<0.001), except for sharpness (p = 0.34). Diagnostic confidence improved from 2.52 to 2.86. Lesion detectability remained comparable before and after correction, with no significant differences in agreement rates (McNemar's p>0.10) or AUCs (DeLong's p>0.06) across all lesion types. Agreement rates ranged from 0.866 to 0.985 in the corrected group against the reference, and AUCs from 0.788 to 0.964. The net reclassification index was 2.66%. Corrected images passed all ACR criteria in phantom testing.

Conclusion: The diffusion model-based algorithm effectively improves image quality and diagnostic confidence without compromising lesion detection, supporting its potential for clinical use in portable brain CT.

Keywords: Diffusion Model; Motion Correction; Portable Brain CT.

Figure 1.

Flowchart for case collection.

Figure 2.

The process of algorithm development and clinical validation. Note that during model training, we used fixed CT as reference standard since it is practically more challenging to find portable CT completely free of image artifacts due to the nature of portable technology.

Figure 3.

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).

Figure 4.

Visual examples of improved lesion detection by motion correction. (a) It illustrates correct reclassification of a positive case (up|event). Motion artifacts obscure the presence of an IPH. The corrected image restores the lesion’s true morphology, which is confirmed in the reference. (b) It illustrates correct reclassification of a negative case (down|non-event). A suspected IPH in the right frontal lobe is seen in the motion image, but it disappears in both the corrected and reference, suggesting it was a false-positive finding induced by minor motion artifacts that was successfully corrected by our algorithm. The reference images in all three examples were scanned on the same day as the portable scan. Image display window is [0,80]HU. The right bottom corner zooms in the region of interest. The reference images in (a) examples were acquired on the same day of the portable scan, while in (b) the reference was acquired one day earlier.

Figure 5.

ROC curves for lesion detection in the motion (blue) and corrected (orange) group. The shaded areas indicate the 95% CI. (Color version of figure is available online.)

Figure 6.

Visual examples of incorrect reclassification by motion correction. (a) shows an example of down|event, while (b) shows an up|non-event case. (a) Both the reference and motion images show a subtle IPH in the temporal horn, which is no longer visible in the corrected image. This may be due to the motion correction algorithm removing the fine details mistaken as motion artifacts. (b) The motion and corrected show parenchymal hemorrhage mostly, the reference (2 days later) shows more intraventricular extension. This suggest there was interval evolution. The mass effect appears less evident in the reference scan, likely due to ventricular decompression, and was therefore scored as “no mass effect” by the readers. The motion image was also scored as “no mass effect,” whereas the corrected image was scored as “mass effect,” likely because improved visualization enhanced reader confidence in identifying it. The reference image in (a) example was acquired on the same day of portable CT scan, while the reference in (b) was scanned two days later.