Deep learning-based bone removal in head and neck computed tomography angiography: a comparative study with
Jie Gao1, Yicun Zhang2, Beibei Shao2
1School of Computer Science, Zhengzhou University of Aeronautics, Zhengzhou, China.
Background:
The adjacent bones and vessels in complex anatomic regions presented diagnostic challenges, especially in head and neck computed tomography angiography (CTA). It requires more accurate removal of bones for clinical routine diagnosis, which couldn't be fully satisfied with conventional bone removal techniques. This study aims to evaluate the performance of a novel deep learning-based bone removal algorithm compared to a conventional method in head and neck CTA, focusing on image quality and radiation dose optimization across two tube voltage settings (100 and 120 kVp).
Methods:
In a single-center randomized controlled trial (RCT) (February to March 2024), 119 consecutive patients [median age, 57 years; interquartile range (IQR), 50-65 years] suspected of cerebrovascular disease underwent head and neck CTA on a dual-source computed tomography (CT) scanner. Patients were randomized to 100 kVp (n=58) or 120 kVp (n=61) groups. Images were processed using a conventional threshold-based algorithm and a convolutional neural network (CNN)-based deep learning algorithm, which was trained on a dataset of 1,014 annotated CTA images. Two blinded radiologists assessed image quality (bone removal effectiveness, vessel branch completeness, whole vessel completeness) with a 5-point Likert scale. Radiation dose was recorded with CT dose index volume (CTDIvol) and dose-length product (DLP). Statistical analysis was performed with Wilcoxon signed-rank tests, Mann-Whitney U tests, and multivariate regression, and P<0.05 was regarded as significant.
Results:
The image quality of deep learning algorithm significantly outperformed that of conventional method across all metrics (P<0.001), with large effect sizes (Cohen's d, 0.886-1.028). Bone removal scores were higher at 100 kVp (median, 4.50; IQR, 4.50-4.50) than those at 120 kVp (median, 4.50; IQR, 3.50-4.50; P=0.002) with the deep learning algorithm, despite lower radiation doses at 100 kVp (CTDIvol, 8.4±0.9 vs. 12.5±1.2 mGy, P<0.001). Age negatively influenced whole vessel completeness (β=-0.0114, P=0.015), while body mass index (BMI) and hypertension showed no effect (P>0.05).
Conclusions:
The CNN-based bone removal algorithm enhances vascular visualization in head and neck CTA, particularly at 100 kVp, offering superior segmentation accuracy even at lower radiation dose. These findings advocate for its integration into clinical workflows to improve cerebrovascular diagnostics.
