Author: Zhaoyang Fan, Eric Nguyen, Dan Ruan, Jiayu Xiao 👨🔬
Affiliation: Department of Radiation Oncology, University of California, Los Angeles, Department of Radiology, University of Southern California, University of Southern California 🌍
Purpose: MR vessel wall imaging (VWI) has been shown to be effective for evaluating intracranial atherosclerosis disease. However, VWI typically also requires an MR angiography (MRA) in the same imaging session to localize blood vessels, which introduces additional time and misalignment risk. We propose a novel MRA-free method for full vessel mapping based on VWI image alone to infer the vessel tree from atlas samples ranked by geometry-relevance criterion with deep metric learning.
Methods: Thirty paired VWI and MRA volumetric images were used in this study, of which 27 pairs were used for training and atlas construction, while 3 pairs were reserved for testing. All images were registered to a common space. Vessel labels were extracted from the MRA volumes, and Dice scores were calculated between all corresponding 2D slice pairs between vessel trees to capture geometric similarity. A Siamese neural network was trained using contrastive loss on pairs of VWI image slices to learn a distance metric in the VWI intensity space that would accurately reflect dissimilarity in the vessel geometry space. During testing, we rank each 3D atlas compared to a target 3D VWI by calculating the mean slice-wise distance between both volumes. Normalized discounted cumulative gain (NDCG) was used to assess the quality of our atlas-ranking process. Vessel inference was then performed by propagating the labels from the closest atlas to the target volume.
Results: Our model achieved a ROC-AUC of 0.88 when evaluated on the test set. Atlas-ranking achieved a mean NDCG@5 of 0.68±0.04. Visual comparison of the inferred vessel trees with ground truth vessel trees shows reasonable alignment.
Conclusion: Our results demonstrate that our model was able to successfully learn a distance metric on the VWI intensity space that captures the underlying geometric vessel tree agreement which can be used for optimal atlas selection.