Project Scope
The long-term goal of this project was to produce software capable of reading magnetic resonance image (MRI) data from any patient and reporting thrombogenesis-related fluid dynamic metrics. The software operates by means of a machine-learning algorithm. The machine-learning algorithm was trained by a sufficient number of CFD analyses to maintain patient-specific accuracy capable of a treatment plan contribution. The importance for professionals in clinics to understand the pathological blood clot formation pathways without fluid dynamic expertise has become increasingly important as Covid-19 patients have experienced a drastically increased incidence of blood clots. Providing a software to model the potential for blood clots will improve the care of future patients by understanding how the clots configure and the risks associated.
Phase I of the effort included the identification and preparation of a single patient-specific anatomy, production of numerous CFD analyses, construction of several machine-learning algorithms, training of these algorithms via the CFD analyses, and analysis of the machine-learning algorithms to determine which is most accurate. Phase II worked toward a production-ready software, by improving the applicability and fidelity of the machine-learning algorithm produced from Phase I. Digital engineering was utilized in Phase II by deploying a neural network to answer the main questions of when the software will be capable of predicting thrombogenesis metrics, and eventually the algorithm will be trusted without its CFD training data.
Rapid Thrombogenesis Prediction in the Carotid Artery
Phase I Effort
This effort included the identification and preparation of a single patient specific anatomy, production of numerous CFD analyses, construction of several machine-learning algorithms, training of these algorithms via the CFD analyses, and analysis of the machine-learning algorithms to determine which is most accurate. The most significant outcome of the Phase I effort was a machine-learning algorithm capable of accurately producing minimum, maximum, and mean time-averaged wall shear stress values and fluid washout percentage for a patient-specific anatomy in seconds rather than the weeks it took the CFD analysis. Other outcomes included code construction of additional machine learning algorithms that can be examined in the future if results dictate, CFD processing script construction useful in future CFD analyses, and CFD temporal and spatial sensitivity studies which, having been completed, will save time and cost as the project expands. The figure shows CFD results representative of the training data used in phase I.
Phase II Effort
Outcomes from Phase II all contribute to the ultimate production of a software product capable of rapid, accurate thrombogenic metric prediction based on MRI data. Phase II improved the applicability of the algorithm to a wider patient cohort by increasing the number of parameters considered for training data production. Phase II efforts improved the algorithm fidelity by introducing wall deformation in the training data. Finally, Phase II efforts hypothesized a minimum number of training data sets required to maintain algorithm accuracy and then test that hypothesis. This outcome was essential to continue applicability expansion as it guides the ultimate timeline for transitioning the machine learning algorithm to software production.
Partners
Institute of Digital Engineering USA
University of North Carolina at Pembroke
Corvid