Automatic Mitral and Tricuspid Valve Flow Quantification using Valve Tracking in 4D Flow MRI

Sammanfattning

In this project a tool for valve flow quantification from 4D flow data for the mitral and tricuspid valves was developed and integrated into the clinical image analysis software Segment. The pre-implemented neural networks MVnet and TVnet were used for tracking the valves in 2CH and 4CH cine MRI images. The positions of the valves were used to create planes in the 4D flow data to measure flow over one cardiac cycle. An algorithm was created to automatically segment valve flow in these planes where flow from other vessels, such as the aorta, is removed while still including potential backflow in the valve. The automatic tracking performed well in adults, but showed reduced performance for the tricuspid valve in some pediatric cases where tracking failures were observed. The automatic segmentation performed well overall with manual correction required in only 5/44 mitral and 7/42 tricuspid cases. The run time in Segment is around 10-15 seconds which provides a short analysis time compared to a solution with a less automated workflow. The implemented mitral and tricuspid valve flow quantification methods were validated on datasets comprising different pathological groups acquired on Siemens (n=37) and Philips (n=7) scanners. In the validation the mitral volume corresponding to the flow through the valve during one cardiac cycle was compared with both 4D and 2D aortic measurements. The tricuspid volume was compared with 4D and 2D pulmonary artery measurements. The validation demonstrated strong correlations and low biases but revealed unexpectedly large variability between all volume measurements, making validation challenging. Comparisons with another established valve tracking tool in 4D flow showed largely comparable results. Encouraging performance was demonstrated, however further investigation and evaluation of the implemented valve flow quantification methods are needed due to remaining variability.