Deep learning assisted image-based cardiovascular disease diagnosis: advancing model explainability and multi-view integration

Abstract

The major aim of the thesis is to investigate the application of deep learning methods in assisting cardiovascular disease diagnosis, non-invasively, based on cardiac images. The disease topics that were explored span from impaired left-ventricular (LV) function, aortic valve (AV) regurgitation, to pulmonary hypertension (PH). From chapter 2 to 5, a non-invasive image-based deep learning approach was adopted towards assistive diagnoses of the respective diseases. Multi-view integration is an important sub-focus of the thesis. From chapter 3 to 5, we gradually verified the benefit brought by using more than one view, establishing a late fusion framework, advancing the fusion module, and finally fusing many noisy and diverse views. Another sub-focus concerns model explainability. From chapter 2, 4, to 5, we first attempted to highlight highly contributing pixels in each individual image with an established method, and gradually moved towards drawing collective conclusions regarding relevant features for a disease.