AI for automatic vertebral motion tracking of fluoroscopic images

For centuries, doctors have been presented with spinal disorders that followed injury or overuse and were therefore thought to be due to loss of spinal stability. In the past, this could not be confirmed, but using a technology called Quantitative Fluoroscopy (QF), which relies on low-dose motion X-ray images captured during motion, it now can.

QF is regarded as the gold standard for spinal motion measurement, but due to lack of readiness for routine use, it is rarely available to scientists, researchers and clinicians (mainly spinal surgeons and physical therapists). In this project we aim to bring Artificial Intelligence (AI) to bear on QF so it can be used more widely in research and patient care.

The reason QF is not in wide use is that it involves the locating and tracking of vertebral motion images by a computer, allowing movement of vertebrae to be measured automatically. Without automation, the process is too time consuming for routine use. At the same time, automatic measurement poses multiple challenges if the images are not very clear and well aligned. Distorted (e.g., due to scoliosis), degraded (e.g., by fat in obese people) or fading (e.g., people with osteoporosis) images currently lead to almost inevitable failure of measurement process - denying the ability to use QF widely for investigating spinal disorders.

We will apply AI methods so that computers can be trained to do both the registration and the tracking of vertebral images at incredible speed, even using images of poor quality. This will allow a huge number of inspections to be performed very quickly, resulting in much more detailed scrutiny and verification and at a much higher speed than a human operator. This will initiate a new era in spine care research.

This project will leverage the QF data which has already been collected over multiple years of prior research, where the participants consented for their data to be used for future research. No additional participants will be recruited, hence no additional X-ray radiation risk will be required for this project.

Low back pain is the greatest cause of disability and is thought to be largely related to spinal mechanics. However, the world of spinal biomechanics research is some way away from being able to provide detailed patient-specific assessments of the mechanics of spinal linkages that could lead to better treatments. The hurdle to be crossed is the provision of an efficient method for measuring the motion of the vertebral linkages and analysing the outputs from dynamic sequences of specific movements for abnormalities.

Prof. Breen and his team have developed such a method using fluoroscopy (low-dose motion X-rays) - and called Quantitative Fluoroscopy (QF). Use of QF in research has seen international growth, largely due to developments in semi-automatic image analysis, enabling processing of large numbers of images and allowing the identification of biomarkers for nonspecific back pain in patients. However, its potential has always been limited by the processing time and limitation to pristine spinal images. This research aims to overcome these barriers.

The team led by Prof. Budka will develop a system using AI and Computer Vision techniques for real-time vertebral detection, identification and tracking, using the existing database of individual in vivo vertebral fluoroscopic images compiled by Prof. Breen. Traditional Computer Vision approaches to object detection rely primarily on object appearance and do not explicitly leverage the hierarchical structure of the scene like the part-whole relationships. Exploiting such structure is key in situations where the image quality is poor and the objects themselves are partially or sometimes even fully occluded.

The methodologies and algorithms we develop will be applied to the analysis of an independent sample of spinal motion sequences obtained at the Royal Hampshire County Hospital as a check on their robustness and universality of application, and indication of any necessary domain adaptation.