The Open Collections website will be undergoing maintenance on Wednesday December 7th from 9pm to 11pm PST. The site may be temporarily unavailable during this time.
UBC Theses and Dissertations
Multi-task learning for leveraging noisy labels in automatic analysis of ultrasound images Mahdavi, Mobina
Supervised machine learning is the standard workflow in training state-of-the-art deep neural networks to automatically analyze, classify, or quantify ultrasound images or videos. However, there are certain challenges regarding the available label set in this context. Since expert annotation is a tedious process, fine-grained and extensive labels are usually not available. The size of data is often limited, and missing labels are an issue. Due to observer variability among different annotators, there is variability in ground truth that manifests as a type of label noise. This thesis aims to investigate the use of multi-task learning to alleviate these issues, in the context of two problems. The first problem is echocardiographic video landmark detection, to quantify the dimensions of the left ventricle of the heart. We first propose a two-headed U-net-shaped convolutional neural network, to detect pairs of inner and outer landmarks on the left ventricle. The model is weakly supervised on the two frames of the video with annotation, with the outer landmarks missing at one of them. Secondly, we propose Differential Learning, which adds a task of ejection fraction comparison to the landmark detection framework, in a Siamese architecture that is trained end-to-end with the main tasks. This auxiliary task is designed to have very low observer noise, by comparing samples that have sufficiently different ejection fractions. We show that this multi-headed model overcomes the issue of missing labels, and Differential Learning improves the results by providing a less noisy training signal. The second problem is biomarker detection, and disease classification in lung ultrasound videos, to detect Covid-19 infection. A multi-headed attention model is first proposed to detect lung biomarkers (A-lines and B-lines) that appear sporadically in lung ultrasound videos, trained on a private and limited dataset with coarse video-level labels. We then propose knowledge transfer to fine-tune this network on the disease classification task in a public lung ultrasound video dataset. We validate this method's ability to overcome limitations in data and label, through ablation studies and comparison to the state-of-the-art. Our proposed attention-scaled explainability method also visualizes the model’s attention to clinically relevant features.
Item Citations and Data
Attribution-NonCommercial-NoDerivatives 4.0 International