Please note: This PhD defence will take place in DC 2314 and online.
Blake VanBerlo, PhD candidate
David R. Cheriton School of Computer Science
Supervisors: Professors Jesse Hoey, Alexander Wong
Point-of-care lung ultrasound is an increasingly important examination in acute care. However, it has yet to be adopted broadly due to a lack of experts and training opportunities. Machine learning-based solutions are being built to address the skills gap; however, labelled training data is difficult to obtain. Self-supervised learning (SSL) has emerged as a method for utilizing unlabelled data to pretrain deep neural networks that extract salient features. In this thesis, we show that SSL improves image classifiers for core tasks in lung ultrasound. We also propose and evaluate novel methods that enhance performance on these tasks.
The first two chapters demonstrate that self-supervised pretraining with contemporary joint embedding SSL methods improves the performance of lung ultrasound classifiers. We show that fewer labelled examples are required to achieve strong performance after SSL pretraining. Further, we observe that models pretrained with SSL exhibit greater generalizability to data originating from external healthcare institutions.
Informed by the knowledge that SSL is beneficial for developing machine learning approaches to lung ultrasound, we propose novel techniques to enhance the efficacy of joint embedding SSL in lung ultrasound. Data sampling strategies in contemporary methods were not designed with ultrasound imaging in mind. We propose two ultrasound-specific approaches to producing semantic pairs for SSL. First, we sample images that are temporally or spatially proximal to each other that originate from the same ultrasound video. The idea is based on the intuition that images that are close to each other will share similar content. Second, we propose ultrasound-specific data augmentation and image preprocessing methods for pretraining. We perform a comprehensive investigation of the efficacy of various data augmentation strategies on different types of lung ultrasound classification tasks. The results indicate the effectiveness of the methods for different types of lung ultrasound classification tasks.
To attend this PhD defence in person, please go to DC 2314. You can also .