Feature extraction and quantification to explore human vasculature

Abstract

Human vasculature plays an important role in maintaining human body function. And vascular diseases are the top leading cause of death around the world. While advanced vascular imaging techniques are available to visualize human vasculature and disease regions, clinically relevant information in medical images (artery structural information, atherosclerotic plaque burden, etc.) remains to be extracted, quantified, validated and explored. Challenges exist preventing the feature extraction and quantification on human vasculature, partly due to the difficulties from the small artery region in the images, complex geometry of arteries, variable signals around vessel boundaries, and limited datasets available for vascular images. In this dissertation, a toolbox of novel algorithms (the Cafe family) was proposed to facilitate 3D vascular image analysis so that comprehensive features of human vasculature can be extracted and quantified automatedly. The endpoint of the analysis is a quantitative vasculature map, which includes three categories of useful imaging features: artery structural features such as artery length, plaque morphometry features such as vessel wall thickness, and vascular disease features such as potential artery segments with atherosclerotic lesions. Three key technical innovations were proposed: 1) Three artery centerline tracing and labeling methods (tracklet refinement, iCafe, AICafe) based on different vascular beds and applications for extracting artery structural features; 2) Y-net and polar segmentation algorithms for extracting plaque morphometry features, including both inner vessel wall (lumen) and outer wall; 3) a domain adaptive classification model for extracting atherosclerotic lesion features. Generating centerlines of arteries (list of 3D points with radii) is the starting point for the quantitative vasculature map construction, as centerlines not only provide artery structural features, but also identify region of interest for further vascular analysis. Three different approaches were proposed for artery tracing based on different vascular beds and purposes of applications. 1) For arteries with relatively straight structures, such as carotid or popliteal arteries, a slice-based artery detection method combined with the tracklet refinement algorithm was proposed to generate centerlines along the luminal centers of arteries. 2) For more challenging arteries with tortuous paths, such as intracranial arteries, a semi-automated image processing technique with graphical user interface (iCafe) was proposed to generate artery centerlines and label anatomical names for arteries with human corrections. 3) When fully automated process is needed, artificial intelligence (AI) techniques were applied to the iCafe workflow, so that the AICafe (AI+iCafe) method trained with processed iCafe results can automate artery tracing and labeling. With artery centerlines and anatomical labels generated using either of the three approaches, a series of artery structural features can be extracted, such as artery length, volume and tortuosity for global or user defined artery groups. These features, as quantitative representations of artery structure and cerebral blood flow, have demonstrated to be useful in various vascular research. After artery centerline generation, artery patches (region of images with artery in the center) can be extracted along the artery centerline for lumen/vessel wall segmentation to further assess luminal structures and atherosclerotic plaques. Traditional manual or semi-automated segmentation methods are time-consuming and only applicable to large straight arteries, limiting its usage in clinical setting or medical research. We propose fully automated lumen/vessel wall segmentation methods (Y-net and Polar models) for delineating lumen and outer wall boundaries based on different Magnetic Resonance (MR) sequences. For bright blood MRI (such as MR angiography) where only lumen areas are visualizable (bright in signal), Y-net (a patch-based lumen segmentation algorithm using convolutional neural network) is used to segment lumen regions from 3D images. For black blood MRI (such as MR vessel wall imaging (VWI)) where both lumen (black in signal) and vessel wall (bright in signal) are visualizable, lumen/vessel wall segmentation is more challenging due to the variable vessel wall signals and flow artifacts. We propose a polar-based method to segment vessel walls in the polar coordinate system so that the lumen/vessel wall boundaries are smoother and more continuous than traditional Cartesian-based methods. In addition, segmentation confidence was available from our model to suggest optional manual checking on challenging slices. The automated lumen/vessel wall segmentation was useful in various vascular research. For example, our FRAPPE tool (a member of Cafe family, designed for popliteal vessel wall analysis) reduced the analysis time for popliteal arteries in a knee MR scan from 3 hours (manual) to 7 minutes (automated), which made large population analysis on lumen/vessel wall feasible. In addition to vascular structure features and plaque morphometry features, identification and classification of possible segments of arteries having vascular disease (for example, atherosclerotic lesions) can provide additional vascular disease assessment features for the quantitative vasculature map. LATTE (another member of Cafe family) was designed for automated lesion identification and classification. In combination with a 2-minute MR VWI imaging sequence (MERGE) and an image quality assessment module, LATTE classifies artery slices along the centerlines into normal arteries, early lesions and advanced lesions, so that patients with vascular diseases can be identified and their artery segments with vascular diseases can be highlighted. One major challenge for robust lesion classification is the domain shift between different MR datasets (signal variations due to different scanners, imaging parameters, coils, etc.), due to which machine learning models developed from a single dataset may not perform robustly when deployed to new datasets. In order to reduce the domain shift, an unsupervised domain adaptation algorithm was applied on lesion classification. Without additional annotations, the CNN adapts its parameters based on the domain irrelevant signals from both the source and target datasets so that the final classification performance can be improved. LATTE was fast in analysis and robust towards various datasets. In addition to extract lesion related features, LATTE can also be a good candidate to be used clinically as a screening tool for vascular diseases. With the Cafe family toolbox loaded with novel algorithms of medical image analysis and machine learning, artery centerlines were identified, lumen/vessel wall regions were segmented, and regions for vascular disease were located and categorized automatedly. The multi-dimensional features extracted and quantified from vascular images formed a quantitative vasculature map which provides comprehensive information for vascular health and helps us better explore the disease pattern of human vasculature.