Joint clustering and component analysis of spatio-temporal shape patterns in myocardial infarction

The Left Ventricle (LV) undergoes remodelling after Myocardial Infarction (MI). In order to quantify the remodelling status, clinicians make use of conventional measures, not fully exploiting the available shape information. To characterize the changes in heart shape and classify heart data as normal or infarcted, we use a hierarchical generative model, which jointly clusters shape point sets from LV in End-Systolic (ED) and End-Systolic (ES) phases, and estimates the probability density function (pdf) of each cluster. We use a Variational Bayes (VB) method to infer the clusters labels, the mean models, and variation modes for the clusters. We also present the results in the supervised setting, where the labels of training data sets are given. Our classification results are evaluated in terms of sensitivity, specificity, and accuracy using 200 LV shapes provided by MICCAI 2015 STACOM LV Statistical Shape Modelling Challenge. Our method successfully classifies the data, achieving a specificity of 0.92 ±0.06 and a sensitivity of 0.96 ±0.07 for the supervised learning approach, and a specificity of 0.83 ±0.03 and a sensitivity of 0.97 ±0.01 for the unsupervised learning approach