- Abstract: The study of cardiovascular diseases requires viable animal models, and small animals have become models of choice due to their significant advantages. Segmentation from cine-MR image has become the gold standard for cardiac function assessment. While many image analysis methods have been developed for clinical studies, similar techniques are generally lacking for preclinical cases. Recent application of neural networks has shown encouraging results in several medical imaging applications. For cardiac cine-MR image segmentation, convolutional neural networks and recurrent neural networks have been successfully used for segmentation of the left ventricle. However, most methods only use a stack of short axis images, which introduces inaccuracy and uncertainty for cardiac function assessment in 3D, especially considering the existence of misalignment between slices. In this paper, we propose an efficient and 3D consistent segmentation method for small animal cardiac MR images taking advantage of a combination of long-axis and short-axis images, by combining convolutional neural networks and the guide-point modelling method. Unlike in most clinical studies, we also focus on training with small datasets, as is common in preclinical studies, and show accurate results with only 12 cine-MR sequences.
- Keywords: segmentation, cardiac MR, cine MR, deep learning, U-Net, small data