This project will open up a new, non-invasive means for assessing cancer. Presently, cancer diagnosis and treatment monitoring using a magnetic resonance imaging (MRI) scanner require the injection of a potentially harmful contrast agent, whereas contrast-free alternatives are too inefficient and imprecise for clinical use. The proposed approach will combine highly efficient imaging strategies with artificial intelligence (AI), to permit fast, contrast-free and robust cancer diagnostics.
During the first two years of the project, we have contributed to the development of new methods that use AI to estimate parameters relevant to cancer diagnosis from MRI images, for both the pancreas and the brain, which are faster and more accurate than conventional approaches. Furthermore, we have made progress towards the development of novel, accelerated MRI imaging strategies, which have the potential to be more robust to common sources of imaging artifacts, or to provide complementary information about the tissue properties.
A branch of AI called “deep learning” involves training a computer algorithm, by analogy referred to as a “neural network”, to detect patterns in data. This training typically requires masses of training data where the information of interest is provided as a known ground truth. Using AI, we have developed a means for generating synthetic MRI data for training neural networks to produce higher quality MRI images from accelerated MRI imaging protocols.
The standard approach to investigating cancer using magnetic resonance imaging (MRI) involves the injection of a potentially harmful contrast agent. Such contrast is contraindicative for patients who are pregnant, breast-feeding, or have kidney problems, and recent studies have also found evidence of long-term contrast deposition in the brain.
Intravoxel incoherent motion (IVIM) MRI is the most promising approach to safe, non-contrast perfusion imaging for cancer diagnosis and treatment monitoring. However, IVIM suffers from high sensitivity to measurement noise and imaging artefacts, and has very long acquisition times, such that currently it is not clinically feasible.
We propose a radical new approach to sampling IVIM data that will drastically reduce acquisition time and ameliorate several artefacts. We will use deep learning to enable efficient data acquisition, robust image reconstruction, and accurate IVIM parameter estimation. Deep learning has never before been applied to MRI in this manner, and it is especially well-suited to IVIM imaging. Successful outcomes promise to herald a new age in MRI pulse sequence development, and to transform current clinical strategies in oncology.