Few Shot Anomaly Detection and Segmentation in Medical Imaging

Deep Learning, a subfield of Machine Learning, has revolutionised the field of medical imaging analysis by automating a wide range of tasks such as disease detection, image reconstruction and the segmentation of lesions and organs. Despite its success, the development and deployment of Deep Learning models in clinical practice is limited by their requirement for large annotated datasets. The process of obtaining annotations requires specific domain expertise which is both tedious and costly. This PhD thesis aims to address this challenge by exploring Deep Learning techniques for medical image analysis in the setting where there is limited training data and annotations. Specifically, it focuses on two core areas; Few Shot Anomaly Detection (FSAD) and Few Shot Segmentation (FSS). Few Shot Anomaly Detection (FSAD) refers to the task of automatically identifying data samples that deviate from normality, in the setting where only few unlabelled examples are available for model training. Siamese networks, traditionally trained in a supervised manner for tasks such as facial recognition are known for their capabilities in the Few Shot setting. This work therefore proposes a Siamese based network, FewSOME for the task of unsupervised FSAD. FewSOME is shown to perform accurate Anomaly Detection having trained on only 'few' unlabelled examples. This analysis demonstrates that FewSOME performs competitively with existing Anomaly Detection techniques that were trained on thousands of images, whilst having significantly lower complexity in terms of training data size and training time. The task of Anomaly Detection has been extensively explored in the medical field for disease detection tasks but less so for other use cases. This work identifies two novel applications of the proposed FewSOME; automated Motion Artefact Detection and continuous disease severity grading. This work demonstrates that the proposed FSAD approach can accurately identify Motion Artefacts in brain MRI, highlighting its potential as a real-time flagging system to prevent repeat scans and improve clinical efficiency. Furthermore, this work shows that FewSOME's anomaly scores on a dataset of knee X-rays exhibit a positive correlation with the severity of knee Osteoarthritis, validating its potential as a continuous grading system. By employing a FSAD approach for both use cases, the requirement for large annotated datasets is eliminated, thus reducing the barriers to clinical implementation. In the second core area of this research, Few Shot Segmentation (FSS), the generalisation capabilities of existing segmentation models in the limited data setting are evaluated. This analysis found that FSS, promptable Zero Shot and supervised approaches can generalise effectively to a new segmentation task on a novel dataset of limited size. Notably, the 3D supervised approach, nnU-Net performs competitively with FSS and promptable Zero Shot models despite the supervised approach being designed for settings with abundant training data. It also achieves this performance with no level of supervision at inference time, while both FSS and promptable Zero Shot models require a sample of few annotated examples or coarse bounding box prompts. The generalisation evaluation was conducted on the DyABD dataset, a novel dataset of abdominal dynamic MRIs, acquired pre-operatively and post-corrective surgery. The MRIs were acquired over time as each patient performed various exercises, thus introducing unique challenges to the task of segmentation due to muscle movement and deformations. This is the first work to develop a fully automatic Deep Learning model for segmenting abdominal muscles in dynamic MRI for patients with abdominal hernias. Additionally, this work contributed to the annotation and pre-processing of the dataset.