Introduction

Segmentation of the left ventricle (LV) and right ventricle (RV) from cardiac MRI is a key step in quantitative analysis such as ejection fraction calculation. Convolutional neural networks (CNN) have led to great progress in achieving fully automatic segmentation; however, one challenge in cardiac MRI segmentation from short-axis images is the different imaging views due to the variability in heart orientation. By its nature, a CNN is shift invariant and partially scale invariant due to the convolution and pooling operations; however, it is sensitive to image rotations due to the rectangular shape of the receptive field so that a trained network may yield sub-optimal results for views that are not well represented in the training dataset. A common method is to apply rotations with a few given angles as training augmentations, but the performance is still limited if the training dataset is small. In this study we propose to train a view-independent network with full-scale random rotations during training and deploy this network with rotation-based testing augmentation to completely eliminate the effect of views.

Methods

In this study, LV and RV segmentation from short-axis cardiac cine images was performed. Although in theory 3D networks can provide through-plane information to improve segmentation accuracy, in practice there can be significant shift among slices, causing worse performance than just relying on in-plane information. Therefore, an optimized 2D U-Net¹ structure was adopted in this study. As dilated convolutions can increase the receptive field without increasing the number of parameters, they were also used during the encoding path and comparisons with normal convolutions were performed. For LV segmentation, the dataset from the 2009 Cardiac MR Left Ventricle Segmentation Challenge² containing 15 subjects for training and 15 for validation was used. Subjects included normal volunteers and myocardial infarction and LV hypertrophy patients. For RV segmentation, the training dataset from the 2012 Right Ventricle Segmentation Challenge³ containing 16 subjects was used. 9 subjects were randomly selected for training and 7 were used for validation. Due to the small sizes of the two datasets, extensive data augmentation including affine and deformable transformations was used during training. To handle the view variability, during training, instead of rotating each image with a few fixed angles, a rotation angle randomly selected from 0 to 360 degree was used in order to train a view-independent network. During testing, each image was rotated 9 times at an interval of 40 degrees and the corresponding output probability maps were rotated back and averaged. In addition to eliminating the variability in views, such test augmentation can also smooth the output contours to be more realistic.

Results

Table 1 shows the Dice scores using the 2D U-Net and 2D dilated U-Net with and without the rotation-based testing augmentation. The testing augmentation improved the mean values and reduced the standard deviations for both U-Net and dilated U-Net structures. Due to the small dataset, no statistical significance was found using a paired t-test. Compared to the original U-Net, the dilated U-Net yielded improved performance. Note that the Dice scores are similar to those of the top performing methods in both challenges. Fig. 1 shows the LV contours on an apical slice with reduced contrast and Fig. 2 shows the RV contours on a mid-ventricular slice. With the proposed rotation-based testing augmentation, the accuracy and smoothness of the contours are improved, especially on difficult slices.

Discussions and Conclusion

One key element in the success of CNN is the large training dataset in order to cover as much variations as possible in real-world applications. Imaging view is a main source of variability in cardiac MRI and is particularly detrimental due to the fact that CNN is not rotation-invariant. Instead of obtaining a large training dataset to cover such variability, we proposed to use a simple rotation-based training and testing augmentation strategy to eliminate such variability. The effectiveness is demonstrated with improved Dice scores and visual contour quality. Furthermore, it can be combined with any network structure and used in any segmentation applications from short-axis cardiac MRI.

Acknowledgements

No acknowledgement found.