I am a PhD student in the WMG Data Science group under Prof Giovanni Montana, working on applying computer vision and generative AI to medical image analysis.

My research focuses on integrating generative AI with medical image analysis to produce intuitive, explainable models for clinicians. I predominantly work with Generative Adversarial Networks (GANs), which are a a class of machine learning models which can create realistic new artificial images or change features in a real image. This technology has broad applications in medical image analysis, such as synthetic dataset augmentation, image super-resolution, de-noising, anomaly detection and modality transformation. Combining generative models with traditional 'black-box' prediction methods allows the features of interest to the model to be visually interpreted, and can show the areas contributing to an AI diagnosis. Improving clinician and patient trust in AI predictions in such ways is key to moving such models from the lab to clinical practice.

Assessing the Performance of Automated Prediction and Ranking of Patient Age from Chest X-rays Against Clinicians; Macpherson et al, MICCAI 2022. ArXiv paper link.

In this paper we present a machine learning model to predict patient age from frontal chest X-rays to ±3 years accuracy, comparing it to the performance of human radiologists. We use a generative AI model to synthetically re-age patient images, highlighting the key anatomical features relevant to age prediction.

 

Patient re-identification from chest radiographs: an interpretable deep metric learning approach and its applications; Macpherson et al 2023, Radiology AI.

We present a model to determine if two frontal chest X-rays come from the same or different patients with high accuracy. By cloning the learned feature representation into a GAN, we visually explore the main anatomical features used to determine patient identity. We show that changes in a patient's identifying features over time can indicate the emergence of abnormalities, giving a potentially useful longitudinal health signal.

 

Development of open-source deep neural networks for comprehensive chest X-ray reading: a retrospective multicenter study in the UK; Dicente, Macpherson et al 2023, The Lancet Digital Health.

We present an open-source model, X-Raydar, to classify frontal chest X-rays across a taxonomy of 37 comprehensive findings, with a public website allowing easy online assessment of user-provided DICOM scans (www.x-raydar.info). We compare the performance of the model against the historical reporting radiologists on a ground truth hand-labelled test set, showing superior or non-inferior performance on the majority of findings.

• 2019-present: PhD student, Warwick University Mathsys CDT. Applications of Generative Adversarial Networks in Medical Imaging, supervised by Prof. Giovanni Montana.
• 2018-2019: MSc - Mathematics of real world systems, Mathsys (Distinction)
• 2008-2016: Portfolio Manager, Pine River Capital Management Hong Kong. Relative value strategies fund manager.
• 2001-2008: Director, Barclays Capital. Convertible bond trader, London and Hong Kong.
• 1997-2001: MSci + MA - Natural Sciences, Cambridge University (1st). Experimental and theoretical physics specialisation.

Generative models, computer vision, metric learning, machine learning in finance.

Contact

Email: matthew.macpherson

@warwick.ac.uk