Supervised Learning of Anatomical Structures Using Demographic and Anthropometric Information

Abstract

A supervised learning approach to predict anatomical structures derived from computed tomography (CT) images using demographic and anthropometric information is proposed. The approach applies a dimensionality reduction technique to a training dataset to learn a low-dimensional manifold representing variation of organ geometry or variation of the CT intensities itself, which computes a mapping between a low-dimensional feature vector and the organ geometry or CT volume. Regression analysis is then applied to determine a regression function between the low-dimensional feature coordinates and external measurements of the subjects such as demographic or anthropometric data. Then for an unseen subject, the low-dimensional feature coordinates are predicted from external measurements using the computed regression function. Subsequently, the organ geometry or the CT volume is estimated from the mapping computed in the dimensionality reduction. As an example case, lung shapes and thoracic CT scans were analyzed based on available demographic parameters (age, gender, race) and anthropometric measurements (height, weight, and chest dimensions). The training dataset consisted of lung shapes represented as a topologically consistent point distribution model (PDM) and CT volumes ($256^{3}\,\mathrm{voxels}, 1.5^{3}\,\mathrm{mm}/\mathrm{voxel}$) of 124 subjects. The prediction error of lung shape of an unknown subject based on 11 independent demographic and anthropometric variables was $10.71 \pm 5.48\,\mathrm{mm}$. Isomap analysis of CT volumes revealed that 95 % of the total variance was explained with 4 dimensions, and the computed mapping clearly captured trends in anatomical variation. This suggested a potential for a direct CT-volume based statistical analysis using dimensionality reduction, which we call a voxel-based statistical atlas. Potential application areas of the proposed approach includes subject-specific ergonomic design in personal protective equipment or population-specific finite-element modeling in biomechanical analysis. Examples also include the use of a predicted patient-specific CT volume as it a prior information for image quality improvement in low dose CT, and optimization of CT scanning protocols.