Deep Learning-Based Versus Threshold-Based Bone Removal for CT Angiography of Head and Neck Arteriovenous Fistulas: A...

Background/Objectives: Bone overlap at the skull base limits CT angiography (CTA) in head and neck arteriovenous fistulas (hAVFs), and threshold-based bone removal fails where bone and contrast attenuation overlap. Whether deep learning-based removal improves the display of the structures that determine endovascular planning is unknown in a DSA-confirmed cohort. We therefore compared the two methods within the same CTA datasets. Methods: In this single-centre retrospective study, 35 patients with DSA-confirmed hAVF underwent head and neck CTA and digital subtraction angiography (DSA) within two weeks of each other, and each CTA dataset was post-processed with both methods on one workstation. Two blinded readers scored the fistula site, feeding artery and draining vein (0 to 2 each) and overall image quality (5-point scale). Scores were compared with paired Wilcoxon signed-rank tests, vessel edge sharpness and post-processing time with paired t tests, and technical visualisation with McNemar's test. Results: Deep learning-based removal displayed the fistula site, feeding artery and draining vein more completely (median total anatomical score 6.00 versus 3.00; p < 0.001 for both readers) and overall image quality was higher (5.00 versus 4.00; p < 0.001). Objective vessel edge sharpness was concordant (316.47 ± 46.68 versus 273.70 ± 45.66 HU/mm; p < 0.001). Technical visualisation was numerically higher (97.1% versus 88.6%) but not significantly so (p = 0.248; three discordant pairs), which makes this endpoint inconclusive rather than negative. Post-processing time was shorter (40.56 ± 1.61 s versus 101.70 ± 4.39 s; mean paired difference 61.14 s, 95% CI 59.53 to 62.74 s; paired t test, p < 0.001). Conclusions: In DSA-confirmed hAVF, deep learning-based bone removal displayed the structures that determine endovascular planning more completely than threshold-based removal and was about one minute faster per case. It does not replace DSA for definitive assessment of fistula architecture, and the apparent advantage at the cavernous sinus is exploratory and requires external validation in larger multicentre cohorts.