3D Regional Shape Analysis of Left Ventricle Using MR Images: Abnormal Myocadium Detection and Classification.

Accurate detection of abnormal myocardium regions is essential for differential diagnosis of cardiovascular disease. However, to achieve this goal by image analysis will significantly increase the burden on radiologists who is already overwhelmed. To ease the time and energy-demanding process and enhance the reproducibility, we proposed a novel framework for automatic abnormal shape detection on left ventricular (LV) using MR images. Our proposed approach utilizes the features obtained by large deformation diffeomorphic metric mapping (LDDMM). To take advantage of 3D structural information, we introduce multilinear principal component analysis (MPCA) in the framework to reduce feature dimensions. Then we combine MPCA with linear discriminant analysis (LDA) to perform differential diagnosis. The performance of proposed framework is evaluated on patients’ images. In the classification of three common cardiovascular diseases, our proposed method outperformed traditional classifiers (Global point signature, Random Forest and XGBoost) with an accuracy of 94%. To further automatically detect the dysfunctional heart regions, we did a comparison on 3D morphology between the diseased subjects and healthy controls and performed an automatic visualization of the abnormal myocadiac regions. In conclusion, our proposed framework reserves the spatial information of the features generated through LDDMM registration and enables the 3D visualization of abnormal regions of LV. With the advance of our method, differential diagnosis is successfully performed on patients with different cardiovascular diseases.