Left ventricular segmentation from cardiac MRI data

Analysis of cardiac magnetic resonance imaging data is essential in diagnosis of many cardiovascular diseases. To assist doctors in manual interpretation, numerous automatic methods are being developed. Recent segmentation algorithms based on deep learning perform the best on homogeneous data, but poorly on the data from different sources. For this reason, modern methods struggle in real world application. Improvement of the effectiveness on the unknown data can be done via several techniques like transfer learning or domain adaptation. In this paper, we proposed a segmentation framework with strong generalisability skills. The pipeline includes domain adaptation with CycleGAN architecture and segmentation of the left ventricular using U-Net 2D convolutional neural network. The method was evaluated on the data set from M& Ms challenge which is formed from data acquisited by several MRI scanners, in different medical centers. Presented method obtained a weighted dice coefficient score of $90.15\pm 6.40$% and weighted Hausdorff Distance of $11.74\pm 13.11 mm$. The results show that applied methods improved generalization of the model and decreased gap between metrics scores for different vendors. Presented method can be classified among other participants of the challenge.