Contactless mid-upper arm circumference estimation from 2D images using deep learning and geometric reconstruction for nutritional screening

Mid-Upper Arm Circumference (MUAC) is widely used for nutritional screening, including among Women of Reproductive Age (WRA). However, manual MUAC measurement may be affected by landmark identification, tape placement, observer technique, and tissue compression. This study evaluates an automated contactless MUAC estimation framework that integrates YOLOv10-based detection, Segment Anything Model (SAM)-based segmentation, deterministic bounding-box-guided ray-casting, and two-view geometric reconstruction from frontal and lateral 2D images. A curated dataset of 144 standardized images from 72 participants was assessed using a 5-fold subject-disjoint cross-validation strategy to reduce identity leakage. The proposed measurement procedure extracted paired mask-chord lengths from frontal and lateral left-arm images and used them as effective axes in Ramanujan’s ellipse perimeter approximation. Across the successfully processed cases, the automated estimates achieved a mean absolute error of 2.51 cm, RMSE of 3.29 cm, MAPE of 9.30%, Pearson correlation of 0.796, and ICC(2,1) of 0.684 with a 95% confidence interval of 0.36–0.83. Bland–Altman analysis showed a systematic positive bias of +1.82 cm, indicating that contactless optical estimation is not directly interchangeable with manual tape measurement without additional bias correction and independent validation. An exploratory modality-specific optical threshold of 25.32 cm produced preliminary alignment with manual classification distributions, but this threshold requires validation in larger independent cohorts. These findings support the technical feasibility of combining detection-guided segmentation and two-view geometric reconstruction for contactless MUAC estimation, while emphasizing that broader clinical generalizability remains unproven .