Abstract:
We present a workflow for producing a statistical shape model (SSM) of the femur with automatically defined regions resembling general anatomic features. Explicitly defined regions enforce correspondence of anatomical features, and allow the shapes of regions to be analysed independently if needed. A training set of manually segmented femur surfaces are partitioned according to Gaussian curvature. Partitioned regions across the training set are then grouped using mean-shift clustering to identify the most stable regions into which surfaces are divided. Reference piecewise parametric meshes are designed for and fitted to each region, and used to train regional SSMs (rSSMs) through fitting?training iterations. Fitted region meshes are assembled into full femur meshes for training a whole femur SSM. Partitioning, clustering and shape modelling results are presented for 41 femurs. In comparison to a non-regional SSM, the rSSM was more efficient and correspondent in its approximation of unseen femurs. We present a workflow for producing a statistical shape model (SSM) of the femur with automatically defined regions resembling general anatomic features. Explicitly defined regions enforce correspondence of anatomical features, and allow the shapes of regions to be analysed independently if needed. A training set of manually segmented femur surfaces are partitioned according to Gaussian curvature. Partitioned regions across the training set are then grouped using mean-shift clustering to identify the most stable regions into which surfaces are divided. Reference piecewise parametric meshes are designed for and fitted to each region, and used to train regional SSMs (rSSMs) through fitting?training iterations. Fitted region meshes are assembled into full femur meshes for training a whole femur SSM. Partitioning, clustering and shape modelling results are presented for 41 femurs. In comparison to a non-regional SSM, the rSSM was more efficient and correspondent in its approximation of unseen femurs.