Open Collections

Ahmadi Amiri, Seyedeh Neda

University of British Columbia

Aortic stenosis (AS) is a valvular cardiac disease that results in restricted motion and calcification of the aortic valve (AV). AS severity is currently assessed by expert cardiologists using Doppler measurements from echocardiography (echo). However, this method limits the assessment of AS to hospitals with access to expert cardiologists and comprehensive echo service. This thesis explores the feasibility of using deep neural network (DNN) for AS detection and severity classification based solely on two-dimensional echocardiographic data. While several machine learning (ML) frameworks have been developed for echo diagnosis and other medical applications, most of them rely on black-box models with low trustworthiness, and they cannot be trained effectively due to the scarcity of training data. However, a model’s explanability and generalizability are essential for clinician adoption. Therefore, a model should be able to capture critical information in both spatial and temporal dimensions of echo videos and provide explanations to support its decision. This thesis proposes frameworks that enhance the state-of-the-art by offering explanability and accurate assessment from echo videos, making ML more practical in echo examination. The first proposed framework aims to compare well-known video models in ML literature for detection and severity assessment of AS. In addition, we leverage semi-supervised learning to fine-tune model weights with unlabeled data. In the second framework, we propose a spatio-temporal architecture that effectively combines both anatomical features and motion of the AV for AS severity classification. Our model can identify the frames that are most informative towards the AS diagnosis and learns phases of the heart cycle without any supervision and frame-level annotations. Furthermore, our method addresses common problems in training deep networks with clinical ultrasound data, such as a low signal-to-noise ratio and frequently uninformative frames. Finally, we address the issue of explanability by incorporating two prototypical layers into existing architectures, enabling interpretable predictions based on the similarity between input and learned prototypes. This approach offers clinically relevant evidence by highlighting markers like calcification and restricted movement of AV leaflets, aiding in more accurate screening examinations.

Text

2023-06-22

Attribution-NonCommercial-NoDerivatives 4.0 International

http://hdl.handle.net/2429/85139

Electrical and Computer Engineering

University of British Columbia

UBCV

http://creativecommons.org/licenses/by-nc-nd/4.0/