Machine learning for outlier detection in medical imaging

Abstract

Outlier detection is an important problem with diverse practical applications. In medical imaging, there are many diagnostic tasks that can be framed as outlier detection. Since pathologies can manifest in so many different ways, the goal is typically to learn from normal, healthy data and identify any deviations. Unfortunately, many outliers in the medical domain can be subtle and specific, making them difficult to detect without labelled examples. This thesis analyzes some of the nuances of medical data and the value of labels in this context. It goes on to propose several strategies for unsupervised learning. More specifically, these methods are designed to learn discriminative features from data of a single class. One approach uses divergent search to continually find different ways to partition the data and thereby accumulates a repertoire of features. The other proposed methods are based on a self-supervised task that distorts normal data to form a contrasting class. A network can then be trained to localize the irregularities and estimate the degree of foreign interference. This basic technique is further enhanced using advanced image editing to create more natural irregularities. Lastly, the same self-supervised task is repurposed for few-shot learning to create a framework for adaptive outlier detection. These proposed methods are able to outperform conventional strategies across a range of datasets including brain MRI, abdominal CT, chest X-ray, and fetal ultrasound data. In particular, these methods excel at detecting more subtle irregularities. This complements existing methods and aims to maximize benefit to clinicians by detecting fine-grained anomalies that can otherwise require intense scrutiny. Note that all approaches to outlier detection must accept some assumptions; these will affect which types of outliers can be detected. As such, these methods aim for broad generalization within the most medically relevant categories. Ultimately, the hope is to support clinicians and to focus their attention and efforts on the data that warrants further analysis.