Author: Rajiv Gupta, Rehab Naeem Khalid, Min Lang, Michael H Lev, Quirin Strotzer, Matthew Tivnan, Maryam Vejdani-Jahromi, Dufan Wu, Siyeop Yoon, Chen Zhennong 👨🔬
Affiliation: Massachusetts General Hospital 🌍
Purpose: Patient motion is a major source of artifacts in portable brain CT due to the slow scanning speed. A diffusion model was developed to reduce these motion artifacts. This work aims to assess the algorithm's impact on image quality and lesion detectability with a human reader study.
Methods: A 3D diffusion model was developed using 100 fixed brain CTs and simulated motion. It was applied to portable brain CTs from 67 patients, each of whom had a reference standard fixed CT obtained within 2 days of the portable CT. Axial images from the following 3 image sets were reviewed: original portable CT (CTpor,ori), AI-corrected portable CT (CTpor,AI), and fixed CT (CTfix). Each image set was scored independently by three neuroradiologists, for the presence of eight lesion types (EDH, IPH, SDH, SAH, IVH, pneumocephalus, mass effect, and hydrocephalus) and four image quality metrics (overall quality, sharpness, artifacts, and confidence), using a 5-point Likert score. The image set type (CTpor,ori/CTpor,AI/CTfix) was blind to the readers. There were at least 2 weeks of washout between scoring different image sets from the same patient.
Results: We established the ground truth for lesion presence by taking the median score on CTfixed and binarization with a threshold of 3. By comparing each observer's score to this ground truth, the AUCs for each lesion type ranged from 0.82 to 0.94 for CTpor,ori, and 0.80 to 0.95 for CTpor,AI. DeLong's test showed no statistical difference for all eight lesion types (p>0.05). Mean image quality scores from CTpor,AI significantly outperformed CTpor,ori by 0.33 to 0.79, using the Wilcoxon signed-rank test for overall quality, artifacts, and confidence (p<0.01), however with similar performance for sharpness (p=0.68).
Conclusion: In this reader study, diffusion-based motion correction for portable brain CT significantly improved both image quality and diagnostic confidence, without reducing diagnostic accuracy.