1st Supervisor: Thomas Booth, King’s College London
2nd Supervisor: Marc Modat, King’s College London
Tertiary Supervisor: Jorge Cardoso
Aim of the PhD Project:
- Patients often have non-specific symptoms e.g. headache or memory loss, which might or might not represent neurological disease
- Because MRI is central to detecting brain abnormalities, patients often undergo work up
- Aim: automatically label brain MRI as normal/abnormal (e.g. tumour/atrophy)
- How: recurrent neural network or deep learning transformer approaches
Project description/background:
The objective of this project is to develop a decision-making tool that identifies abnormalities on magnetic resonance imaging (MRI) brain scans using deep learning.
This will be applied to patient groups where there are non-specific symptoms such as headache or memory loss.
A clinically validated decision-making tool that identifies abnormalities on brain MRI scans does not exist. This project will meet this clinical need by providing a deep learning model that can automate the identification of abnormalities immediately in ‘real-world’ conditions.
This is important as over 330,000 patients are waiting more than 30 days for their MRI reports in the UK. This number is forecast to increase, as there is a greater demand for MRI than there is availability of radiologists to report these scans. UK-specific workforce shortages in clinical radiology are negatively impacting patient care by delaying diagnosis; a similar picture is seen globally.
Immediate triage of a brain MRI into normal or abnormal potentially allows early intervention to improve short- and long-term clinical outcomes. In general, detecting an abnormality or illness early would result in lower costs for the healthcare system because less specialist medicine and fewer hours of treatment are needed for the patient to recover.
The objectives are:
- Build a classification model to determine normal and abnormal brain scans
- Clean datasets for patient groups where there are non-specific symptoms such as headache or memory loss.
- Validation on non-research grade data of the classification model identifying abnormalities on brain MRI scans to assess possible clinical use for prioritisation and/or triage of true negatives (i.e. normal scans).
References
- Dixon S, “Diagnostic imaging dataset annual statistical release 2017/18”
- The Royal College of Radiologists, “clinical radiology UK workforce census 2019 report”
- Brennan S, “National radiology review launched over ‘significant backlog’ fears,”https://www.hsj.co.uk/.. national-radiology-review-launched-over-significant-backlog-fears
- NHS England, “Transforming imaging services in England a national strategy for imagingnetworks” https://improvement.nhs.uk
- McKinney SM et al. International evaluation of an AI system for breast cancer screening.Nature 2020:57789–94.
- FDA-NIH Biomarker Working Group. In: BEST. Food and Drug Administration (US) 2016.
- Wood et al. Automated triaging of head MRI examinations using convolutional neuralnetworks PMLR MIDL https://arxiv.org/abs/2106.08176v1 2021
- Morris Z et al. findings on brain magnetic resonance imaging: systematic review and metaanalysis.BMJ 2009;339:b3016.
- Booth TC et al. Management of incidental findings during imaging research in “healthy”volunteers: current UK practice. Br J Radiol. 2012;85:11–21.
- Lu S. Detection of abnormal brain MRI via improved AlexNet and ELM optimized by chaoticbat algorithm. Neural Com & Appl 2020 https://doi.org/10.1007/s00521-020-05082-4
- Talo M et al. Convolutional neural networks for multi-class brain disease detection usingMRI images. Computerized Medical Imaging and Graphics 2019;78:101673
- Rauschecker AM et al. Artificial intelligence system approaching neuroradiologist-leveldifferential diagnosis accuracy at brain MRI. Radiology, 295:626–637, 2020.
- Rudie JD et al. Subspecialty-level deep gray matter differential diagnoses with deeplearning and bayesian networks on clinical brain MRI. Radiology AI, 2:e190146 2020
- Biomind Available: https://www.biomind.ai [Accessed 10 feb 2020].
- Baur RG et al. Steganomaly: Inhibiting cyclegan steganography for unsupervised anomalydetection in brain MRI. MICCAI 2020, 718–727, Springer 2020
- Han C et al. MADGAN: unsupervised medical anomaly detection GAN using multipleadjacent brain MRI slice reconstruction, 2020. https://arxiv.org/abs/2007.13559
- Baur C et al. Scale-space autoencoders for unsupervised anomaly segmentation in brainMRI. MICCAI 2020, 552–561, Springer, 2020.
- Chen X et al. Unsupervised lesion detection via image restoration with a normative prior.MIA 2020 https://doi.org/10.1016/j.media.2020.101713
- You S et al.. Unsupervised lesion detection via image restoration with a normative prior.PMLR MIDL 102, 540–556, 2019. URL http://proceedings.mlr.press/
- Kobayashi K et al. Unsupervised brain abnormality detection using high fidelity imagereconstruction networks, 2020. https://arxiv.org/abs/2005.1257320
- Zimmerer D et al. Context-encoding variational autoencoder for unsupervised anomalydetection, 2018. https://arxiv.org/abs/1812.05941
- Zimmerer D et al. Unsupervised anomaly localization using variational auto-encoders.MICCAI 2019, 289–297, Springer, 2019.
- Baur C et al. Deep autoencoding models for unsupervised anomaly segmentation in brainMR images. MICCAI Brainlesion Workshop, 161–169. Springer, 2018.
- Wood D et al. Deep learning to automate the labelling of head MRI datasets for computervision applications. European Radiology online first July 2021
- Wood D et al. Labelling imaging datasets: validation study MICCAI arXiv:2007.04226v3
- Wood D et al. Automated labelling using an attention model for radiology reports of MRIscans (ALARM). PMLR MIDL 2020 arxiv.org/abs/2002.06588v1
- Zaharchuk G. Fellow in a box: combining AI and domain knowledge