A four-zone surface classification of the subclavian–axillary artery and zone estimation from body-surface depth images using semantic segmentation

Abstract

Axillary hemorrhage is difficult to control before hospital arrival because effective external compression depends on locating a deep arterial target beneath variable shoulder anatomy. We defined four bilateral body-surface zones for the subclavian-axillary artery and evaluated their estimation from computed tomography angiography (CTA)-derived depth images. Of 456 retrospectively reviewed CTA scans, 230 were technically eligible and 228 cases with paired depth images and labels underwent fixed patient-level five-fold out-of-fold (OOF) evaluation. The selected U-Net used equally weighted cross-entropy and multiclass Dice losses with depth-safe augmentation; comparators were standard U-Net, HRNetV2-W18, nnU-Net v2 2D, and nearest-template affine registration. The selected U-Net achieved overall OOF Dice 0.7397 ± 0.0586, with zone-wise Dice 0.5606, 0.7574, 0.7771, and 0.8637. Relative to standard U-Net, mean paired Dice differences were 0.0122 overall (95% confidence interval [CI], 0.0088–0.0156) and 0.0503 for Zone 1 (95% CI, 0.0391–0.0617). Median Zone 1 centroid distance was 10.86 pixels, and 201/228 cases (88.2%) had an intersection over union (IoU) of at least 0.25. These internal CTA-derived simulations support computational feasibility; prospective external validation with real color-and-depth (RGB-D) image acquisition and clinical outcome assessment remains necessary.