An untreated stroke event would be destructive for a human deep space exploration mission. Increased cerebrovascular disease risk has been documented after prolonged exposures to ionizing radiations on Earth. Astronauts on deep space exploration missions will be exposed to galactic cosmic rays and solar particles for 30 months, which will lead to accelerated vascular injury likely increasing their risk of stroke, which is exacerbated by the negative effects of microgravity on cerebrovascular autoregulation.
On Earth, acute strokes can be successfully treated with anticoagulant or thrombolytic drugs if the event is rapidly diagnosed and the type of stroke (ischemic versus hemorrhagic) is rapidly identified with Computer Tomography (CT) or Magnetic Resonance Imaging (MRI) brain imaging. However, these brain imaging capabilities do not exist in space and alternative robust means of diagnosing and classifying stroke are needed. Due to the homology between retinal and cerebral vessels, and the ease with which retinal images can be acquired non-invasively, retinal images have been studied as a marker for cerebrovascular events. We propose to use a combination of color fundus photos and Optical Coherence Tomography Angiography (OCT-A) images to identify stroke events and stroke type, effectively acting as a proxy for brain imaging. These imaging modalities are non-invasive and deployable in currently existing technologies for deep space missions. We will adapt our automated interpretable image-based deep learning algorithm to identify stroke and stroke type from retinal vascular images, enabling an automated life-saving tool usable on a deep space exploration mission.
Original project aims/objectives:
Aim 1: Build a terrestrial acute stroke dataset retinal images and stroke clinical outcomes.
Aim 2: Develop and validate an interpretable deep learning model to identify the presence of stroke and stroke type from color fundus photos.
Aim 3: Development and validation of interpretable deep learning model to identify presence of stroke and stroke type from OCT-A (and fundus image combination).
Project Highlights and key findings:
• Camera acquisition setup completed, data acquisition team trained and 128 subjects recruited.
• All images have been externally graded by researchers blinded to the stroke condition of the subjects.
• In our dataset, it is possible to detect acute ischemic stroke with relatively high-performance using OCT-A images (AUC 0.87 [CI 0.78-0.99]). Age does not appear to be a confounder.
• 10 macular microvasculature density variables per retina are enough to achieve these performance.
• Fractal dimension on fundus images and OCT-A are associated with acute ischemic stroke but they are worse predictors than macular microvasculature density variables.
• Multi-modality integration (fundus+OCT-A) does not increase the predictive performance in our dataset.
• State-of-art self-supervised deep learning algorithms do not increase the predictive performance even when pre-trained on other two external OCT-A datasets (without the stroke information).
• Developed a live web application to generate the stroke predictions from the 10 macular microvasculature density variables per eye. This would allow other researchers to use and further validate our model. The web app is available at a web address, as source code and as Docker container.
• Developed a new deep learning model to extract a synthetic OCT-A image from fundus images. This has the potential of having better vessel segmentations without the need of being trained on manual segmentation.
• Released a dataset of OCT-A, fundus images, our synthetic OCT-A and vessel segmentations to allow other researchers to use and further refine our synthetic OCT-A model.
• Retina vasculature embeddings tested on non-acute stroke data. We found a significant predictive association with our retina vasculature imaging biomarker and stroke subjects. The model developed outperformed two state-of-the-art deep neural networks. All the embeddings have been submitted to UK-Biobank database for sharing with other investigators.
• Identified multiple challenges in the data acquisition that can inform future hardware development.
Impact of key findings on hypotheses, technology requirements, objectives and specific aims of the original proposal:
• Our results support our initial hypothesis that it is possible to detect acute stroke from retina images.
• We did not need to use symmetric differences between the two retinas as initially planned to train our model.
• We were not able to recruit enough hemorrhagic stroke patients to test our ability to distinguish between ischemic and hemorrhagic stroke. This is partly due to the COVID pandemic did not allow us to start the data acquisition as initially planned, partly to the fact that hemorrhagic stroke subjects have more disability which leads to issues in focusing onto the camera and/or suffer from droopy eyelids, which does not allow successful imaging.
• There is the need of better portable or robot-arm mounted OCT-A cameras to image bedridden subjects.