Use Of Two-stage And Dual-decoder Convolutional U-net Ensembles For Vessel And Plaque Segmentation In Carotid Artery Ultrasounds
Lauren A. Huntress, MD1, Mieyan Xie, Ph.D.2, William E. Beckerman, MD1, Saum Rahimi, MD1, Usman W. Roshan, Ph.D.2, Justin W. Ady, MD1.
1Rutgers-RWJMS, New Brunswick, NJ, USA, 2Rutgers-NJIT, Newark, NJ, USA.
Objective: Carotid ultrasound is a screening modality used to aid in the prevention of ischemic stroke. B-mode ultrasound can identify the vessel wall, lumen, plaque burden, and degree of stenosis, but requires qualified technicians and physicians to perform and interpret the study. Our initial experience with automated deep learning techniques for independent image evaluation has been previously described; albeit limited to vessel lumen identification. Herein, we expand our previous work with use of a dual decoder U-net and two-stage cascaded U-net model to independently segment vessel lumen and identify plaque burden on carotid artery B-mode ultrasounds.
Methods: B-mode ultrasounds from 226 patients (500 images) of the carotid artery with varying degrees of stenosis (<50%, 50-70%, >80%) were obtained. Using previously manually segmented data of vessel lumen and plaque burden within the internal carotid artery (ICA), external carotid artery (ECA), and common carotid artery (CCA), Pytorch was used to develop a convolutional U-net system with 90% accuracy for plaque segmentation. A 10-fold cross validation study was performed to measure segmentation accuracy via the Dice coefficient and a composite confidence score was calculated using the mean U-net output. Accuracy of vessel and plaque segmentation was compared between the two-stage cascaded U-net, dual-decoder U-net, and basic U-net models (Fig. 1).
Results: When provided raw B-mode images, the dual-decoder convolutional U-net model outperformed the basic convolutional U-net with an accuracy of plaque and vessel lumen segmentation of 68.8% (vs. 66.8%), while the two-stage model had an accuracy of 65.1%. With manual input of the vessel, accuracy of the two-stage model rose to 81.7%. At varying confidence intervals (0.5, 0.6, 0.7, 0.8), the dual decoder U-net identified vessel lumen and plaque with an accuracy of (73.2%, 75.2%, 79.5%, and 87.3%) when compared to the basic U-net (72.8%, 75.3%, 79.2%, 85.9%), respectively.
Conclusion: Deep learning continues to be a valuable and underutilized analytical modality in vascular surgery. Our study demonstrates the potential of dual-decoder and two-stage cascaded U-nets and the feasibility of automated structural interpretation of carotid artery ultrasounds.
Back to 2021 Karmody Posters