Deep Learning-Based Intracranial Vessel Wall Segmentation: Good DSC with Biased NWI?
H Zhou1*, J Xiao2, Z Fan23, D Ruan14, (1) Department of Bioengineering, UCLA, Los Angeles, CA, (2) Cedars-Sinai Medical Center, Los Angeles, (3) Keck School of Medicine of USC, Los Angeles, CA, (4) School of Medicine, UCLA, Los Angeles, CA
Presentations
SU-IePD-TRACK 3-7 (Sunday, 7/25/2021) 12:30 PM - 1:00 PM [Eastern Time (GMT-4)]
Purpose: Vessel wall (VW) contouring is important in cardiovascular disease diagnosis. Deep learning methods have been applied to MR VW images to delineate the lumen and the VW, with good common segmentation objectives, in Dice similarity coefficients (DSC). On the other hand, to support quantitative clinical judgement, further measures need to be assessed and characterized. This study provides an in-depth investigation on the behavior of a popular clinical metric – the normalized wall index (NWI), which is defined by the VW area divided by the whole vessel area, on the cross-sectional slices.
Methods: We investigated the combination of (a) different network structures, including 2D UNet, 2.5D UNet, and 2.5D UNet++, all with good DSC performance 0.9163±0.0522, 0.7452±0.1046; 0.9080±0.0641, 0.7521±0.1006; and 0.9116±0.0723, 0.7758±0.0957 for the lumen and the VW, respectively; (b) two objective functions: DSC, and DSC plus Hausdorff distance loss; and (c) different training sample reweighting schemes, including random, uniform, importance sampling, and matched distribution between training and testing to establish performance upper-bound. We examined closely the impact on NWI prediction and its variation along the vessel, to generate hypothesis for the cause of such bias and guide further quantitative radiomic study.
Results: The results showed persistent systematic bias and large variation on the testing NWI, regardless of network structure, objective function, or training set sampling scheme. The NWI bias ranged from 0.0592±0.0237 to 0.0975±0.0425.
Conclusion: NWI may not be sufficient as the sole quantitative basis for plaque burden judgement. Its ratio form could be more sensitive to mislabeling. We hypothesized (1) it may be beneficial to optimize the clinical endpoints directly as the segmentation objective; (2) the seemly small variation in contour labeling in training labels could have a large impact on derived quantitative omics features and their associated confidence. A better scheme to homogenize manual labels is required.
Funding Support, Disclosures, and Conflict of Interest: This work is supported in part by NIH/NHLBI R01 HL147355.
ePosters
Keywords
Cerebral Vasculature, MRI, Blood Vessels
Taxonomy
IM/TH- image Segmentation: MRI
Contact Email