Estimating Subclavian-Axillary Artery Zones for Potential Compression Strategies in Axillary Hemorrhage Control via Depth Image-Based Semantic Segmentation (Preprint)

BACKGROUND Junctional hemorrhage, particularly in the axillary region, presents a major challenge in trauma care. Early and effective external compression are crucial life-saving maneuvers; however, the potential compression sites for various techniques remain unclear, and the effectiveness of these techniques may vary depending on the region and type of hemorrhage. OBJECTIVE This study aims to improve the identification of potential compression sites for axillary hemorrhage control in prehospital scenarios. METHODS The classification of subclavian‒axillary artery zones 1–4 was determined on the basis of arterial branching patterns, anatomical landmarks, and body‒surface structures. A total of 230 axilla CTA datasets were collected and randomly divided into training datasets (n=184) and validation datasets (n=46). The two‒dimensional depth images were derived from the three‒dimensional CTA array of the body surface via a dimensionality reduction method. The landmark points for zones 1–4 were annotated on the CTA slices and then projected onto the depth images to create labeled images. We developed a modified U‒Net model with a coordinate attention block and compared its accuracy in zone estimation with that of the standard U‒Net. The Intersection over Union (IoU), precision, recall, and Dice coefficient were analyzed. RESULTS Compared with the standard model, the modified U‒Net improved the mean IoU, precision, recall, and Dice coefficient in zone 1 by 0.03, 0.01, 0.09, and 0.03, respectively. Among the modified U‒Net model, zone 4 yielded the highest accuracy, whereas zone 1 yielded the lowest accuracy. The median Dice coefficients for subclavian‒axillary artery zones 1–4 were 0.58 (0.52‒0.62), 0.81 (0.74‒0.84), 0.83 (0.78‒0.85), and 0.88 (0.86‒0.91), respectively. Among the 46 validation datasets, 5 datasets presented Dice coefficients exceeding 0.59, 0.81, 0.83, and 0.88 for Zones 1–4, respectively. CONCLUSIONS This study proposed a novel four-zone anatomical framework of the subclavian–axillary artery and demonstrated the feasibility of using deep learning–based semantic segmentation on depth images to localize potential compression regions with high accuracy. The region-specific differences in segmentation performance underscore the need for tailored hemorrhage control strategies based on vascular complexity and surface accessibility. These findings provide critical anatomical and technical groundwork for developing intelligent external compression systems and augmented reality (AR)–assisted visualization tools, thereby enhancing the precision and timeliness of axillary hemorrhage control in prehospital trauma care. CLINICALTRIAL Not applicable (retrospective imaging study; no clinical trial registration).