Abstract
Cardiac ultrasound imaging, also known as echocardiography, is one of the main tools to diagnose heart disease, due to its availability, affordability, ability to provide real-time results, and absence of ionizing radiation. However, echocardiography is challenging and requires a lot of practice to master, both in terms of scanning and in terms of interpreting the images. There is a growing shortage of clinicians in echocardiography, which is expected to worsen in the current decade. Automated analysis of cardiac ultrasound images can help clinicians diagnose heart disease more efficiently and guide non-experts.
In echocardiography, many time-consuming measurements are performed on different views of the heart to provide quantitative information for diagnosis and follow-up of patients. Deep learning and artificial intelligence (AI) tools can play a crucial role in these echocardiography challenges by enabling fast, reproducible and accurate interpretation of cardiac ultrasound images.
Despite advances in AI, its adoption in echocardiography faces challenges such as limited data, varying image quality, and lack of ground truth and expert consensus, which affects generalisability and reliability. The overall goal of this work was to investigate techniques to make deep learning tools for echocardiography more robust. Specifically, the work provides contributions to dealing with three different challenges. The first contribution is a method for detecting failing cases of deep learning based segmentation models. The study demonstrates that failures in segmentation models can be identified using a secondary model specifically designed to prevent anatomical errors. The second contribution is a method for automatically estimating the image quality of different regions of the heart. Different methods were tested and the results showed that higher quality images correlate with tighter limits of agreement between automated and manual clinical measurements. Finally, the last contribution provides a technique to deal with limited available annotated data for segmentation models using a generative diffusion model. The work shows how generative augmentations can improve the robustness and generalisability of segmentation models.