Deep learning for vessel segmentation and flow analysis to identify clusters associated with adverse outcomes in a fontan patient registry
Abstract
Abstract We introduce a deep learning framework comprising two models for automated segmentation (DCS) and large-scale deep temporal clustering (DTC) within a registry of single ventricle patients. The DCS model performs simultaneous classification and segmentation of velocity-encoded phase-contrast magnetic resonance (PCMR) data for five individual blood vessels, the left and right pulmonary arteries, aorta, superior vena cava, and inferior vena cava. Trained, validated and tested on 260 cardiac MRI exams (each containing 5 PCMR scans), it demonstrated a median Dice score of 0.91 on 50 unseen test exams. Integrated into a fully automated pipeline, the DCS model processed over 4500 registry exams without manual intervention, reaching 98% classification accuracy and 90% segmentation accuracy in cases with all five vessels present. Flow curves obtained from successful segmentations were used to train the DTC model, which performs deep temporal clustering to uncover unique flow patterns. Survival analysis showed that these groups were statistically correlated to increased risk of mortality or transplantation and to liver disease, highlighting the clinical relevance of the proposed framework.
Article Details
Authors (37)
Tina Yao
Nicole St. Clair
Madeline Gong
Gabriel F. Miller
Michael Quail
Shahin Moledina
Adam L. Dorfman
Mark A. Fogel
Rajesh Krishnamurthy
Christopher Z. Lam
Joshua D. Robinson
Timothy C. Slesnick
Justin Weigand
Jennifer A. Steeden
Rahul H. Rathod
Vivek Muthurangu
Manish Aggarwal
Tarek Alsaied
Ashish Doshi
Matthew D. Files
Sanjeet Hegde
Andrew Hoyer
Tiffanie Johnson
Yue-Hin Loke
Alison L. Marsden
Laura J. Olivieri
Francesca Raimondi
Preeti Ramachandran
Pierangelo Renella
Markus S. Renno
Garg Ruchira
Amee Shah
Jonathan H. Soslow
Vanderbilt University Medical Center, Nashville
Jeremy Steele
Kenan W.D. Stern
Bijoy Thattaliyath
Aswathy Vaikom House