Thorax shape reconstruction from limited CT-digitized palpable landmarks using statistical shape modeling

This study aimed to develop and evaluate algorithms for reconstructing subject-specific thorax geometries from a limited set of skin landmarks, typically collected in motion analysis studies, using statistical shape models (SSMs). A dataset of DICOM CT scans from 76 individuals was used to generate thorax geometries, comprising the sternum, ribs, and vertebrae C6 to T10, and to collect five bone landmarks and their corresponding skin landmarks. Two thorax SSMs were developed from the CT-derived bone geometries, differing in whether skin landmarks were embedded within the SSM. Based on these models, two reconstruction strategies were implemented using an optimization framework: (i) a bone-landmark-based approach in which multiple linear regression was used to map skin landmarks to bone landmarks (SSM-BL), which served as optimization targets, and (ii) a skin-landmark-embedded approach that used skin landmarks directly as optimization targets (SSM-SL). Skin-to-bone mapping accuracy was quantified using mean absolute error (MAE) and root-mean-square error (RMSE). Thorax reconstruction accuracy was assessed using point-to-point surface errors, reported using MAE and RMSE. Skin-to-bone regression models achieved MAE and RMSE below 5 mm for most landmark coordinates. No statistically significant differences were observed between reconstruction strategies. Thorax reconstruction errors ranged from 8.26 ± 1.96 mm to 8.36 ± 1.70 mm (MAE) and from 9.10 ± 1.12 mm to 9.32 ± 1.49 mm (RMSE), depending on the number of retained principal components. The proposed SSM-based methods enable thorax reconstruction from sparse data, capturing global thorax morphology. However, local anatomical details were less accurately represented.