Real-Time Automated Quality Control for Liver MRI

Principal Supervisors: Prof. Vicente Grau (Institute of Biomedical Engineering, Oxford) and Dr Luca Biasiolli (Radcliffe Dept of Medicine, Oxford University)

Co-Supervisor: Dr Matthew Robson (Perspectum Diagnostics)

Background

The recent advent of large-scale MRI studies, such as UK Biobank [1], and the use of quantitative analysis have stressed the need for automated quality control (QC). The current standard approach is visual quality review, usually performed by multiple human observers to assess variability and reach consensus. This method is inherently subjective and qualitative, therefore not suitable for large-scale trials producing big data, as it would be very time-consuming, expensive, and potentially lead to inconsistent assessments among different observers.

The literature discussing medical image quality (IQ) assessment is sparse and mostly concerned with IQ metrics borrowed from the evaluation of natural images [2]. This means that it is focused on measuring the perceptual visual quality of whole images, i.e. finding global metrics that correlate with human judgement, rather than on quantitative and objective tools to detect the presence of image artefacts at a local level and specific to a certain imaging modality, e.g. MRI.

To our knowledge, the only study trying to address MRI artefacts detection proposed a method based on intensity thresholding on the background of structural brain images [3]. That approach has several limitations and is not generalisable to other anatomies, as it is based on assumptions that are specific to brain MRI and can only detect artefacts that are propagating into the image background (as variations of the noise distribution).

To conclude, it is worth pointing out that automated IQ assessment of in-vivo MRI data is a completely different problem from that of detecting MRI artefacts caused by hardware malfunction, which is part of the routine quality assurance performed on images acquired by scanning standardized phantoms.

Aims and objectives

Currently, a critical limitation for offline generation and analysis of quantitative maps is the inconsistent quality of the acquired MRI data, as operators cannot rely on any objective QC at scan time. We aim to develop and validate a QC tool for liver MRI that can be installed on the scanner for real-time feedback to radiographers, enabling them to rectify the problems before the patient leaves the scanner. It may also find application offline to enforce IQ consistency in multi-centre trials, i.e. to check that data were acquired according to the prescribed protocol and to the required standards, so that they will be analysable.

In the proposed project, we will investigate a novel QC approach for liver MRI that is objective and quantitative, thus highly reproducible. The aim is to focus on the IQ of the specific anatomy and on the imaging artefacts that are corrupting it. The automated algorithm will extract information from images and metadata, and fuse them to detect issues with the MRI acquisition, e.g. slice location, signal-to-noise ratio (SNR), blurring and ghosting due to patient motion. The IQ evaluation will also indicate the likelihood of success for automated quantitative analysis of the liver.

In a simplified version, similar to a real-time traffic light system, the software will classify the quality of MRI scans as ‘good’, ‘review’ (borderline quality that requires the operator’s attention), or ‘poor’ quality (an objective reason to discard data and, if possible, repeat acquisition).

Methods

We will apply state-of-the-art Computer Vision and Machine Learning (ML) techniques to MRI and develop methods to assess IQ of clinical MRI scans automatically without the need of any user interaction. Our strategy will be to combine specific information on MRI sequences extracted from DICOM headers with local image features describing the quality of the region of interest, and with more conventional measures of the global IQ (e.g. SNR).

We will design and implement the algorithms to segment regions of interest and extract a set of high-level image features of the target anatomy that describe different aspects of the local image quality, similarly to the measures of vessel edge sharpness that we have used for carotid arteries [4] and to the methods we have developed for aortic cine scans and briefly presented in the section on preliminary data. The approach will be focused on artefacts degrading the quality of liver MRI.

The development of the algorithms for image feature extraction and the optimal choice of ML technique will be performed on training datasets containing different IQ examples. The automated QC results will then be tested on separate datasets and evaluated against the gold standard IQ scores from visual review performed by multiple human observers, as described in the section on preliminary data.

The training and testing datasets will be representative subsets of liver MRI data selected from the UK Biobank cohort, which is large (more than 18,000 subjects scanned to date, and 100,000 in total) and well characterised. The UK Biobank acquisition protocol is quite difficult to run and might be rolled out to other sites, hence it is pertinent to this approach.

Proposed plan

In the first year, the DPhil student will 1) study the origin of imaging artefacts corrupting liver MRI data acquired by different sequences, 2) extract relevant sequence information from DICOM and 3) experiment with different metrics for image sharpness, SNR and artefacts detection. This work will be presented at an international conference (e.g. ISMRM).

In the second year, the student will 4) start developing the algorithms for automated segmentation, local feature detection and IQ description and, optionally, explore their applicability to other anatomies. He/she will also 5) compare and select the most suitable ML technique for the task of learning the IQ feature values from the training data, and then 6) test the real-world performance on a separate dataset. This work will be presented at a conference on biomedical image analysis (e.g. MICCAI, ISBI).

In the third year, the student will 7) complete the development of algorithms and 8) the training/testing experiments for Machine Learning. At the end of the DPhil, the student’s goal will be to have a fully automated QC pipeline for liver MRI data. This work will lead to the publication of a full article in an international journal (e.g. MRM, JMRI).

Deliverables and Impact

This project will deliver a set of methods that can be used at scan time to support radiographers by giving them immediate feedback or offline to assess IQ automatically and enforce its consistency.

Healthcare companies analysing MRI data post-hoc (e.g. Perspectum Diagnostics) would benefit from the methods that will be developed in this project, as acquisitions with sub-optimal IQ can be detected while the patient is still in the scanner and repeated immediately, avoiding the need for a second visit which inevitably adds considerable time and expense. MRI scanner manufacturers (e.g. Siemens) might be interested to install the algorithms directly on their machines to give immediate feedback to radiographers and indicate when it is necessary to repeat acquisition in real time. Pharmaceutical companies and Imaging Core Labs running multi-centre MRI-based clinical trials could use our tools to monitor data quality in objective terms and enforce IQ consistency across different MRI centres. A re-scan can be requested and the patient can be called back in realistic time frames to avoid missing data points.

References

1. Petersen SE, Matthews PM, Bamberg F, Bluemke DA, Francis JM, Friedrich MG, et al. Imaging in population science: cardiovascular magnetic resonance in 100,000 participants of UK Biobank - rationale, challenges and approaches. Journal of Cardiovascular Magnetic Resonance. 2013;15:46.

2. Chow LS, Paramesran R. Review of medical image quality assessment. Biomedical Signal Processing and Control. 2016;27:145–54.

3. Mortamet B, Bernstein MA, Jack CR, Gunter JL, Ward C, Britson PJ, et al. Automatic quality assessment in structural brain magnetic resonance imaging. Magn Reson Med. 2009;62:365–72.

4. Biasiolli L, Lindsay AC, Choudhury RP, Robson MD. Loss of fine structure and edge sharpness in fastspinecho carotid wall imaging: measurements and comparison with multiplespinecho in normal and atherosclerotic subjects. Journal of Magnetic Resonance Imaging. 2011;33:1136–43.