The research objectives of this project are formulated as part of four key challenges:
1) Development of intelligent algorithms that can learn from each other and exchange information in order to solve complex image analysis tasks;
2) Leveraging large-scale population data and linking images with non-imaging data such as demographics, lifestyle, genetics and disease to construct powerful statistical models;
3) Utilising these models to inform novel approaches for abnormality detection to find subtle signs of disease in medical scans;
4) Obtaining a better understanding on how black box machine learning derives at decisions to gain the trust of doctors, patients and policy makers for use of AI in clinical applications.
The project team has made significant progress on all four challenges, and published more than 50 scientific papers in high impact journals, international conferences and workshops. We have developed and demonstrated several new approaches for multi-modal, multi-task learning leveraging knowledge across different imaging domains and different clinical applications. We have shown that image segmentation can benefit by learning from multi-modal data. We have also demonstrated that solving multiple tasks jointly can improve accuracy and efficiency of image segmentation models. The team has developed new methodology for domain generalisation making machine learning models more robust against changes in the input data which is important for clinical deployment.
A major achievement is the development of a novel approach to causal generative AI. We have developed a novel causal AI approach, deep structural causal models, for high-fidelity image generation under causal assumptions. Deep structural causal models enable, for the first time, high-resolution and plausible counterfactual image generation which as applications in robust machine learning, bias mitigation, and fairness. We have also developed new algorithms for fully automatic, highly accurate image segmentation based on deep learning. Specifically, we have demonstrated the clinical utility of our Brain Lesion Analysis and Segmentation Tool for Computed Tomography (BLAST-CT) for the quantitative assessment of traumatic brain injuries. BLAST-CT uses deep convolutional neural networks to accurately detect, identify and segment different types of bleedings in the brain which provides important information for deciding on treatment strategies and patient management in the setting of emergency and intensive care. The results of our multicentre validation study provide strong evidence for the clinical value of automated image segmentation algorithms.
In order to build trust in AI for clinical use, we need to better understand how complex machine learning models such as deep neural networks work under different settings and under which conditions they may fail. We have developed new methodology for automated quality control and prediction of failure cases that can be used during deployment. In particular, we have proposed new approaches for estimating the performance of AI models in the absence of ground truth. This is critical methodology for AI deployment, allowing us to predict when a model works and when it may fail. Another output of our research is a comprehensive stress testing approach to assess the robustness of image classification models. Stress testing can reveal weak points of a model, facilitating the improvement and auditing of AI models.
