MIRA
Research type
Research Study
Full title
Next Generation Machine Intelligence for Medical Image Representation and Analysis
IRAS ID
233249
Contact name
Ben Glocker
Contact email
Sponsor organisation
Imperial College London
Duration of Study in the UK
5 years, 0 months, 0 days
Research summary
Summary of Research
This project is devoted to develop a new generation of intelligent algorithms for automatically analysing and interpreting medical scans to support doctors when making critical decisions in diagnosis, therapy and intervention. Key to the project is its unique access to some of the largest and most comprehensive imaging databases combined with world-leading expertise in machine learning and medical imaging. An overarching objective is to harvest information from population data to construct what will be the most advanced statistical models of anatomy. Linking these models with demographics, lifestyle, genetics and disease allows probing of genetic and environmental determinants related to specific anatomical and pathological phenotypes across organs. This will provide insights into complex population diseases, and enables a novel approach to abnormality detection that aims to automatically find subtle signs of pathology in new medical scans.Summary of Results
The project has led to a number of advances that push the boundaries of the state-of-the-art in machine intelligence for medical image analysis.We have developed novel methodologies to address key challenges in medical imaging including new learning strategies to leverage diverse, heterogeneous and multi-modal data which has been shown to improve the state-of-the-art in semantic segmentation and detection of pathology. We have demonstrated the benefit of multi-task learning and devised new methodology to exploit unlabelled data via so called semi-supervised learning. We have shown that it is possible and beneficial to learn simultaneously from multiple imaging modalities such as CT and MRI. We have developed a new algorithm for detecting various types of lesions in CT, including very small ones that are clinically highly relevant. Our algorithm surpasses the performance of previous methods by a significant margin. We have also demonstrated that it is possible to estimate the quality of automatically derived predictions which is important for gaining the trust in black box machine learning in clinical settings.
A significant contribution of the project relates to the introduction of a causal perspective on key aspects in machine learning for medical imaging, namely scarcity of high-quality annotated data and mismatch between the development dataset and the target environment. We presented arguments for the importance of taking the causal story behind the data into account when designing machine learning models which can help to identify and avoid issues arising from dataset shift and sample selection bias. We believe that this work will substantially contribute to a novel research direction of causal representation learning in medical imaging. This contribution has the potential to pave the way for a major breakthrough in developing more robust, reliable and trustworthy AI in medical imaging which is the ultimate goal that we want to achieve by the end of the project.
REC name
London - Chelsea Research Ethics Committee
REC reference
18/HRA/0229
Date of REC Opinion
26 Sep 2017
REC opinion
Favourable Opinion