MA-Shape: Modality Adaptation Shape Regression for Left Ventricle Segmentation on Mixed MR and CT Images

Abstract

Left ventricle (LV) segmentation is essential to clinical quantification and diagnosis of cardiac images. While most existing LV segmentation methods focus on cardiac images of single modality or multi-modality, few have been devoted to images of mixed-modality. By Mixed-Modality, we mean that different modalities exist in the database, while for every subject, there is only one modality. In this paper, we propose a newly invented LV segmentation method from mixed-modality images: modality adaptation shape regression (MA-Shape). Compared to single-modality or multi-modality methods, the proposed MA-Shape can 1) be applied to images of new modality during the test phase, which improves the generalization of the learned methods, and 2) take advantage of existing samples of different modalities, which alleviates the high demand for multi-modality data. To achieve this, we propose a modality adaptation module to enhance the shape consistency between the MR and CT, and therefore improve the generalization of the model learned in one modality to new modalities. The experiments on a dataset with MR sequences of 145 subjects and CT scans of 96 subjects validate that the proposed MA-shape can achieve excellent performance by learning common shape information from images of mixed modality and improve the cross-modality generalization of shape regression model learned on images of one modality. These advantages not only provide an efficient way of utilizing mix-modalities data during model learning but also enables an effective and flexible way of applying automated cardiac function assessment in clinical practice.