Ultrasound enables radiation-free longitudinal monitoring of scoliosis, but rib shadowing and speckle noise often obscure vertebral structures. Current deep-learning methods present results in terms of localisation accuracy, without directly measuring anatomical completeness. We introduce a vertebra-level completeness model that includes a YOLO-based detection framework and an explicit representation of completeness, the Vertebra Presence Matrix (VPM). The VPM provides visibility into detections across 17 ordinal vertebral levels (T1–T12, L1–L5), allowing us to measure completeness across anatomy rather than just detections. Thoracolumbar ultrasound scans were annotated and divided into train/test sets using a patient-wise split to avoid data leakage. Four model variants were evaluated, including full-spine and vertebra-centric crop representations with single-class and 17-class detection heads. The full-spine detector was less stable in regions of high anatomical variability, such as the upper thoracic and lower lumbar spine. Crops of individual vertebrae were more stable under partial fields of view. The 17-class crop model achieved an mAP50 of 0.929 and a scan-level completeness score of 0.74 using the VPM. These results demonstrate that vertebral completeness can be explicitly quantified and integrated with localisation-based metrics for completeness-aware automated scoliosis evaluation.
Dana et al. (2026) studied this question.