Damage detection on Psoriatic Arthritis X-rays

Psoriatic Arthritis diagnosis

Assessment of joint damage via X-rays is an essential part of treating and diagnosing psoriatic arthritis. This task is traditionally undertaken by rheumatologists, which is both time consuming and expensive. Improving the assessment of X-rays has huge potential for advancing precision medicine in arthritis treatment including the potential of improving diagnosis and prognosis, understanding which treatments can help prevent damage to joints and improving trial design for novel therapeutic agents.

This project was originally conceived of and developed through the PhD project of Adwaye Rambojun in the EPSRC Centre for Doctoral Training SAMBa. Since the completion of his PhD, Adwaye has moved onto an NHS-funded one-year postdoc under the supervision of Campbell, Shardlow and Tillett. As IMI fellowships they are leverage opportunities together with Adwaye to extend the longevity of the project, realising its full potential impact.

X-rays are scored for three types of damage, namely joint-space narrowing, osteoproliferation, and bone erosion (left-to-right above). Each type of damage is assigned a score and this is done for each joint in the hand. We are developing machine-learning techniques to identify bones and complete this task automatically by fitting a nonlinear shape model and applying classification algorithms.

Machine learning

Data collection and annotation is done by our team of rheumatologists from the Royal United Hospital using our in-house user-interface.

The collected data is used to train the statistical shape model. The shape model is defined by a Gaussian Process Latent Variable Model and can generate new bone shapes and measure uncertainties in the automatic assessment. This model parametrises shapes via a nonlinear mapping from a learned latent space; examples are shown below of bone shapes (left) generated from a four-dimensional latent space (centre and right).

The shape model is then used to identify bones in the hand. We are able to accurately identify bones in the index finger.


We focus on three main aspects of the project which brings together computer science, mathematics and machine learning.
User Interface Design: We are developing a scoring and annotation tool with our models plugged-in that will allow rheumatologists to collate and analyse scores. This work is done in collaboration with MSc student Nicolas Porshke.

Geometric Shape Priors: Bone outlines are closed curves. Rather than modelling point clouds as is normally done, we are developing statistical shape models for continuous objects that respect the underlying geometry.

Deep Learning and Computer Vision: Statistical shape models are able to measure geometric shape changes, but not texture changes due to swelling and inflammation. We are investigating both discriminative models to act as score regressors as well as generative models to increase our understanding of the texture information that characterises PsA damage.

The medical images on this web pages are taken from:

[1] Tillett W et al J Rheum 2016 Feb 43(2):367-70 DOI: 10.3899/jrheum.150114 (top image), [2] Oxford Textbook of Psoriatic Arthritis, Chapter 16 Plain Radiography Tillett W and McHugh N: DOI: 10.1093/med/9780198737582.001.0001 (others).