AI-based chest circumference measurement from a photograph in children: a YOLO11-pose keypoint detection study

Accurate subxiphoid chest circumference assessment is required for the clinical evaluation and treatment planning of children with microtia. Conventional tape measurement requires an in-person visit and is prone to measurement-site error when performed by untrained caregivers. We developed a non-contact AI approach that employs a YOLO11-pose model to detect bilateral chest keypoints and a Ramanujan ellipse model to estimate subxiphoid chest circumference from a single frontal photograph, with the lateral view used solely for verification of chest depth. A single-center observational diagnostic-accuracy study of 500 children with microtia (250 boys, 250 girls) was conducted. Reference subxiphoid chest circumference was measured by trained personnel; AI prediction used YOLO11-pose keypoint detection with Ramanujan ellipse geometric modeling (chest depth = width × 0.75). The cohort was split into a development set (400 cases) and a locked test set (100 cases). Performance was assessed by mean absolute error (MAE), root mean squared error (RMSE), Pearson correlation, and Bland–Altman agreement, with bootstrap 95% confidence intervals. In the locked test set (n = 100), MAE was 0.938 cm (95% CI 0.799–1.089), RMSE was 1.193 cm (95% CI 1.029–1.364), median absolute error was 0.80 cm, and mean bias was − 0.127 cm (95% CI − 0.359 to 0.108). Predictions were within 2, 3, and 5 cm of the reference in 93.0%, 99.0%, and 100% of cases. Pearson correlation was 0.935 (R 2 = 0.856). MAE was 0.897 cm in the normal-weight subgroup and 1.025 cm in the overweight/obesity subgroup. At a 60-cm threshold, sensitivity was 87.3%, specificity was 100.0%, and overall accuracy was 90.0%. A YOLO11-pose and Ramanujan ellipse pipeline can achieve high accuracy for non-contact subxiphoid chest circumference estimation in children with microtia under standardized imaging conditions, and mitigates measurement-site error by non-medical users through automated landmark detection. Because all images were acquired under controlled conditions, the reported performance may not directly reflect real-world in-home use. Performance was slightly lower in the overweight/obesity subgroup, warranting further validation in larger cohorts.