Improving Automated Lung Ultrasound Interpretation with Self-Supervised Learning

Abstract

Lung ultrasound (LU) is an increasingly important point-of-care examination in acute healthcare settings. In addition to exhibiting comparable accuracy to conventional imaging modalities such as radiography or computed tomography, ultrasound provides safety from radiation and enhanced portability at a reduced cost. Despite its purported benefits, LU has yet to be widely adopted due to a lack of trained experts and education programs. In response, machine learning algorithms have surfaced to alleviate the skills gap by providing automated interpretation. However, the training of machine learning algorithms requires vast amounts of manually annotated examples - that is, images labelled by an expert for whether or not they exhibit particular findings. Given that there are few experts qualified to provide quality annotations, there is a need to explore alternative machine learning techniques for this unique problem.

Self-supervised learning (SSL) has emerged as a class of methods to train machine learning algorithms to extract salient features from data without relying on labels. These so-called “pretrained” algorithms constitute better starting points when unlabelled data is abundant but labelled examples are scarce or challenging to acquire, as is the case for LU. Therefore, the purpose of this thesis is to propose and evaluate novel techniques tailored to lung ultrasound that enhance performance on core interpretation tasks, such as detecting lung sliding and pleural effusion.

The thesis consists of four studies. The first study explores the efficacy of SSL techniques in brightness mode LU interpretation tasks and proposes a multi-task framework that reuses a single pretrained algorithm to efficiently interpret LU images. The results demonstrate that SSL improves performance on multiple classification tasks. In addition, experiments show that pretrained algorithms amplify generalizability to external healthcare centres. The second study introduces SSL methods to motion mode ultrasound and proposes data augmentation techniques specific to motion mode. We observe that pretrained algorithms achieve the greatest performance on local and external test data for the challenging task of lung sliding classification. The third study proposes a LU-specific technique for SSL that involves sampling pairs of images from the same ultrasound video that are temporally or spatially proximal to each other, based on the intuition that such pairs of images share similar content. The results indicate that, with appropriate parameter assignments, this sampling strategy improves performance of pretrained algorithms on multiple tasks. Lastly, the fourth study proposes ultrasound-specific data augmentation and image preprocessing methods for SSL. The results underscore the value of ultrasound-specific image preprocessing in SSL. A comprehensive evaluation finds that ultrasound-specific data augmentation yields the best performance on a diagnostic task, and that techniques based on cropping attain top performance on object classification tasks.

Overall, the findings of this thesis demonstrate that SSL improves the performance of machine learning algorithms for LU, especially on LU images originating from external healthcare centres. Novel SSL methods for LU are established as a key ingredient for producing algorithms that are effective for multiple LU interpretation tasks.