Prediction of growth based on shape analysis of atherosclerotic calcifications from lateral X-ray images

We present a method for prediction of atherosclerotic growth based on a training set of 229 2D manually annotated baseline and corresponding follow-up calcifications from lateral X-ray images over an 8 year period. The prediction uses affine shape analysis based on singular value decomposition where non-rigid shapes are modeled as projections of rigid high-dimensional shapes. The SVD prediction was compared to growth based on dilation and predictive conditional PCA. The SVD model yields the largest Jaccard score indicating that the SVD model captures the most of the growth. We also found that small circular shapes grow the most which is likely due to the higher growth potential in smaller calcifications. Furthermore, we are able to predict the biological risk factors better by a joint shape and biology model that suggests a relationship between the shapes and the risk factors.