Clinically trusted Artificial Intelligence and medical image analysis for monitoring inflammatory arthritis

Rheumatoid arthritis (RA) and psoriatic arthritis (PsA) are the two most prevalent forms of inflammatory arthritis that cause autoimmune-induced joint inflammation leading to structural damage, pain and significant disability in patients. Current clinical diagnosis and monitoring of the diseases rely on plain radiographs of hands, wrists, and feet for damage assessment.

Several radiographic scoring systems have been proposed and adopted in clinical research and trials for damage quantification. However, their application is limited by the complexity of manual scoring, inter-observer variability and failure to describe detailed variations in disease manifestation, making it difficult to quantify treatment effects and disease progression.

As a time-consuming process, scoring is rarely performed in clinical diagnosis and monitoring of the disease progression.

With advances in Artificial Intelligence (AI), several automated radiographic diagnosis, grading and scoring frameworks have been proposed for RA, demonstrating promising performance. Nevertheless, the intrinsic issues with existing scoring systems have not been addressed, and most of the models use methods that provide limited interpretability or explainability. In addition, no established AI-based radiographic scoring approaches have been proposed for PsA.

This project aims to develop novel automated radiographic quantification schemes for RA and PsA which could provide finer details of the diseases by adopting interpretable and explainable Deep Learning (DL) techniques. A range of conventional machine learning and DL methods based on convolutional neural networks will be experimented with. We plan to employ the concept of similarity ranking that directly compares the anatomical structure in images to propose a more interpretable model for damage quantification.

To provide explainability, post-hoc explanation methods such as feature weighting and visualisation of learned representations or models will be utilised as the baseline. Self-explainable model structures such as prototype variational encoders which learn the prototypes that may be linked to disease stages in the feature space and their projections in the input space will be explored as well.

The developed damage assessment method could then be deployed on existing hand and feet X-ray datasets linked to electronic health records from RA or PsA patients to study disease trajectories. Subgroup analysis will be performed using clinical data to identify subtypes of disease progression and variations in treatment response. The performance of the proposed methods will be also validated using data from the retrospective clinical trials in the hope of generating novel discoveries.

The proposed studies will establish new explainable automated quantification schemes for RA and PsA that could be applied to grant greater insight into the manifestation of the diseases in clinical settings and potential treatment or personal features that affect their progression in time. The project will be collaborative in its nature including collaboration with Oxford Psoriatic Arthritis Centre and Royal United Hospitals Bath.

The project falls within the EPSRC Healthcare Technologies research theme and the Medical Imaging and AI Technologies research areas. It will lay the foundation for the development of clinically relevant RA and PsA evaluation tools to facilitate treatment decisions in the clinic and treatment effect assessments in clinical trials.