Probabilistic deep learning approaches in medical imaging

点击翻译按钮获取中文摘要

In medicine today, the use of imaging data plays an indispensable role across all aspects of care and spans most medical domains. The effective analysis of this data is paramount, requiring accurate, reliable, and efficient tools. Artificial Intelligence (AI) has emerged as a transformative solution, extensively applied across diverse imaging modalities and diseases. AI has notably enhanced the precision and efficiency of image analysis while reducing the burden on clinicians. However, a challenge persists in the form of AI models struggling to generalize across disparate data sources, such as different hospitals or imaging devices, and adapting to different diagnostic tasks. One promising avenue to address these challenges is the application of statistical and probabilistic approaches to AI models. These approaches can enhance model robustness and reliability, but have yet to be fully explored, particularly within the realm of medical imaging. This research gap forms the basis of my DPhil project. My project will focus on the development and evaluation of probabilistic deep learning tools tailored to medical imaging, using methods from both frequentist and Bayesian statistics. This includes building novel models which are robust to out-of-distribution data, can be trusted, and from which causal, correlative, and confounding effects can be distinguished. A particular focus will be placed on the development of models which can associate a confidence estimate or distribution to each of their predictions. To date, most classical deep learning models only provide a point-estimate of their prediction, and "don't know what they don't know", which can be particularly problematic when they are presented with images from a different distribution to those with which they were trained. Quantifying uncertainty is key to models being trusted by clinicians, especially in decision-sensitive contexts such as healthcare. Despite this, existing medical imaging models which quantify uncertainty, such as Bayesian neural networks or Monte Carlo dropout methods, often incur significantly higher computational costs, as they involve calculating a distribution over each of the models' weights and training a series of networks with different activations for each layer respectively. This project will therefore aim to develop such models which can reliably estimate their uncertainty while maintaining prediction accuracy and low computational cost. I will primarily use multimodal Positron Emission Tomography and Computed Tomography (PET/CT) data of patients with tumours in order to develop and test these models. This type of data would particularly benefit from uncertainty-aware models as there is extensive inter-scanner variability as well as variability in the interpretation of the scans by clinicians. This would benefit the wider deep learning research community, as code would be open-sourced and methods shared, as well as the clinic, by providing safer and more trustworthy methods. For instance, this would be important in the use of PET/CT tumour segmentation to guide radiotherapy, as having a map of the uncertainty across the predicted tumour area would avoid targeting of any potentially healthy areas, eg at the margins, which are notoriously harder to segment. The company GE Healthcare will be involved in the project as the industrial partner and will also help provide curated dataset(s) that I can work with. This project aligns with EPSRC's strategies and research areas. Specifically, this project falls within the following EPSRC research areas: - Artificial intelligence technologies - Image and vision computing - Medical imaging - Statistics and applied probability.