L. Stahnke. CNN-based left ventricle segmentation in transesophageal ultrasound images. Philips Research; Institute of Biomedical Engineering, Karlsruhe Institute of Technology (KIT). Bachelorarbeit. 2021
According to the World Health Organization (WHO), cardiovascular diseases (CVDs) are the leading cause of death globally. In clinical practice, cardiac ultrasound (US) imaging is widely used by cardiologists to determine clinically relevant parameters such as ejection fraction (EF) or the thickness of the myocardium for the diagnosis of CVDs. Measuring these parameters manually is tedious and error-prone. Thus, in this thesis, a deep learning approach for left ventricle (LV) segmentation in three-dimensional (3D) transesophageal echocardiography (TEE) images is developed and the effect of incorporation of prior anatom- ical information on the segmentation result is investigated. This thesis is one of the first studies to demonstrate TEE image segmentation with the U-Net convolutional neural network architecture. Therefore, the hyperparameters of the segmenta- tion pipeline are optimized by using a random search. The main contribution of this thesis is the evaluation of the effect of global image rotations on the segmentation results. Hence, it is shown that a normalization of the scan angle (a global rotation around the z-axis) of TEE images can improve the segmentation results significantly for images with a scan angle larger than 60°. Secondly, it is shown that the LV Dice Similarity Coefficient (DSC) can be significantly improved if the images are rotated based on prior knowledge about anatomical structures. The network’s performance is thereby robust with respect to imprecisely defined axes if the standard deviation of the angles at which the rotation matrix is tilted is less than 15°. Furthermore, a convolutional neural network is trained on images that are cropped around a given region of interest. It is shown that for some parameter sets, inference time was 1.5 times faster without a decrease in DSC, but for other settings, the cropping of images affected quality of LV segmentation negatively. The best experimental setup in this thesis achieved a DSC of 0.927, Hausdorff Distance (HD) of 0.115, and Normalized Surface Distance (NSD) of 0.969 for the left ventricular cavity on the test set. Thereby, the network was trained on 184 manually annotated and 982 automatically annotated images of 84 patients.