Anomaly Detection and Characterisation with Few-Shot Machine Learning

Anomaly detection and image reconstruction are new and exciting frontiers of machine learning, employing the most sophisticated and efficient algorithms available in order to solve real and pressing issues. These types of solutions are incredibly active areas of research across a multitude of fields and subject areas, from detecting credit card fraud to reducing noise in captured images or video. In recent years, medical imaging has been a large sector looking to employ and tailor these methods. One of the most prominent of these specialist uses is in X-ray imaging. Post image reconstruction in this area has been shown to be sufficiently sound, allowing for reduction in scanning time and radioactive dose administered to patients, both of which are crucial in resource constrained settings or in cases of unstable or at risk individuals. When looking to use medical data in learning algorithms, obtaining labels is non trivial and requires scarce and expensive experts. For this reason unsupervised learning systems are much more likely to result in practical and useful implementations. This proposed project would look at researching and developing a novel unsupervised anomaly detection algorithm for use in the X-ray imaging sector where an anomaly could be fraudulent scans or atypical bone structure. Studies similar to this have been conducted on hand and chest images, both concluding with positive results. The project proposed here differs from these by use of image reconstruction techniques in aiding to detect these anomalies as well as the potential use of feature engineering, looking to model bone density and additional spatial information. There already exists multiple theoretical strategies in image reconstruction, all of which would be explored on a fundamental level in in order to develop a new algorithm suitable for use. Algorithms such as generative adversarial networks (GANs) and variational auto-encoders have been found to produce reasonable results and may provide an interesting starting point for research. The proposed feature engineering in this project is much more ambitious, however would ideally set the road for future development in this area. If possible, these additional features could be used in other aspects of X-ray imaging, such as density tagged data in scans where changes could be tracked over time, potentially aiding the diagnosis of conditions such as osteoarthritis and osteoporosis. The additional spatial information may include degree of bone separation, which could aid in the search for other bone related diseases. Feature extraction of bone density has been briefly explored using traditional machine learning methods and some deep convolutional neural networks (CNNs), however has not yet been extensively explored with other deep learning techniques or in conjunction with image reconstruction or anomaly detection. This project is intended to be ambitious from the start, where identifying how these areas of research may be brought together and utilised would be the primary focus. From here, it is likely that the project would narrow it's span and delve deeper into the development of novel deep learning algorithms in one of these discussed sub-areas.