The Tooploox-affiliated research team has delivered the first example of measuring pulmonary artery blood pressure using CT scans instead of a surgical procedure.
Pulmonary Hypertension is a dangerous and incurable disease that affects the arteries of the lungs and the right side of the heart. In one form of the disease, called pulmonary arterial hypertension, the lung’s blood vessels are narrowed, blocked or otherwise damaged, forcing the heart to pump blood with more strength, as well as reducing oxygen supply to the full organism.
The disease itself is impossible to cure, yet there are treatments available that can improve quality of life and prolong life expectancy.
What this type of hypertension has in common with other types of hypertension is the fact that it is a silent assassin, in that it shows few or even no symptoms for a long time, leaving little to no room for prevention. Yet where typical hypertension can be detected with a blood pressure meter at home, pulmonary hypertension requires Right Heart Catheterization.
What is Right Heart Catheterization?
Right Heart Catheterization is an invasive diagnostic test requiring a skilled surgeon to use a probe inserted into the heart via the femoral, brachial, or axillary artery. The procedure is costly and invasive, making the diagnosis difficult and troublesome.
Troublesome, challenging, and expensive? Sounds like a perfect (yet difficult) task for Artificial Intelligence.
Surviving the pressure
The idea behind the research was to develop a system capable of detecting the tell-tale pressure from non-invasive diagnosis tools. To do so, it was necessary to combine data from cardiac computer tomography (CT), physics-based model parameters, Magnetic Resonance Imaging (MRI), and the patient’s demographic data.
All of this data could be relatively easy to handle for neural networks – the challenge was in combining them. Moreover, some of these (including MRI and CT) were image-based data, while others (demographic or physics-based model parameters) were tabular data, which are extremely challenging to combine. It was basically an industry standard to use neural networks to either handle images or tabular data, but never together at once.
Or rather – never before.
TabMixer
To overcome this challenge, the Tooploox research team has introduced TabMixer, a system that allows a single neural network to read both tabular and image data simultaneously without losing context. The system is, therefore, capable of processing an MRI scan while “remembering” that the image shows, for example, an overweight adult man and analyzing changes in blood flow with physical model data.
As such, the system’s output is not an idiosyncratic measurement delivered by the neural network simply for the sake of the measurement but rather an established unit and truly actionable information.
The result
According to the team’s best knowledge, the Tooploox-affiliated work has provided the first way to detect mean pulmonary pressure (mPAP) from an MRI scan. The system not only worked on super-clean data but also on noisy data gathered from real medical cases on various models of diagnostic machines.
Overall, the solution has shown excellent robustness in analyzing low quality of data, combined with state-of-the-art accuracy, fully comparable with the results of Right Heart Catheterization.
Other use cases
Combining the abilities of tabular data analysis and image recognition can be applied in multiple use cases, including:
- Environmental responsibility – the system may analyze a number of cars, identify them to later check engine type and estimate greenhouse gas emissions.
- E-commerce – the system may recognize items in a warehouse using CCTV cameras and combine this information with the warehouse management system to spot inconsistencies or mistakes.
Summary
The research was delivered by a team consisting of Michał K. Grzeszczyk, Przemysław Korzeniowski, Samer Alabed, Andrew J. Swift, Tomasz Trzciński and Arkadiusz Sitek, of the Sano Centre for Computational Medicine, University of Sheffield, Warsaw University of Technology, IDEAS NCBR, Tooploox, and Massachusetts General Hospital, Harvard Medical School.
The full text of the paper can be found on Arxiv.
