Enhanced clinical decisions via image quality transfer

Lead Research Organisation: University College London
Department Name: Computer Science


The potential availability of current bespoke and rare high-end MRI scanners provides the opportunity of improved diagnoses for multitude of human pathologies. However, diffusion of this advanced technology to more standard clinical settings is hindered by the cost of constructing and maintaining such expensive hardware, in addition to market forces, which reduce the speed of adoption.

Our approach is to harness the power of deep learning to learn a mapping between acquisitions from mundane, to state-of-the-art scanners. We first gather a set of subjects to be scanned under both quality scanners. We then use supervised learning on paired cubic patches, where the input, output are patches from respective mundane and high-quality acquisitions. The trained model can then be deployed to improve the quality of unseen mundane scans, where the predicted patches are stitched together.

In previous work, we have shown the benefits of our approach in both developed [1,2] and developing [3] countries. In both paradigms, we have shown clear visual improvement in image and contrast enhancement [1,3] and better quality images to diagnose epileptic lesions [2].

The approach in [1] uses publicly available dataset (HCP) as a proxy for high-quality acquisitions and we create low-quality images with a downsampling operation. We first plan to use the entirety of the HCP cohort (>1000 subjects), allowing us to deploy more sophisticated and complex models. Furthermore, we aim to incorporate personal information such as subject age, sex, and handedness into the predictive model. Finally, we intend to improve the downsampling technique in [3], so that low-quality downsampled images more realistically correspond to low-quality scans. Providing the user the option to control contrast and quality in downsampled images will allow us to simulate a variety of low-quality acquisitions, augmenting the training set and improving model generalisability.

A similar problem is where we aim to harmonize data across different scanner manufacturers and quality settings. When a group of subjects are able to travel to different hospitals to obtain additional acquisitions, we demonstrated the effectiveness of a novel multi-task learning approach [2]. However, paired data is often not available. Therefore we aim to use representation theory in machine learning to map each brain to a common feature space, with a soon-to-be-available large dataset (>1000 subjects).

We hope to deploy methods in both approaches towards an extension of [3], where the aim is to improve the quality of scans of Nigerian children to better identify epileptic lesions. For this project, I am collaborating with both colleagues at UCL and in University Hospital Ibadan.

[1] S. B. Blumberg et al. Deeper Image Quality Transfer : Training Low-Memory Neural Networks for 3D Images, MICCAI 2018.
[2] S. B. Blumberg et al. Multi-Stage Prediction Networks for Data Harmonization, MICCAI 2019.
[3] H. Lin, M. Figini, R. Tanno, S. B. Blumberg et al. Deep Learning for Low-Field to High-Field MR : Image Quality Transfer with Probabilistic Decimation Simulator, MICCAI MLMIR 2019.


10 25 50

Studentship Projects

Project Reference Relationship Related To Start End Student Name
EP/R512400/1 01/10/2017 30/09/2021
1971787 Studentship EP/R512400/1 25/09/2016 30/09/2021 Stefano Blumberg