Image‐ versus histogram‐based considerations in semantic segmentation of pulmonary hyperpolarized gas images

Abstract

Purpose To characterize the differences between histogram‐based and image‐based algorithms for segmentation of hyperpolarized gas lung images. Methods Four previously published histogram‐based segmentation algorithms (ie, linear binning, hierarchical k‐means, fuzzy spatial c‐means, and a Gaussian mixture model with a Markov random field prior) and an image‐based convolutional neural network were used to segment 2 simulated data sets derived from a public (n = 29 subjects) and a retrospective collection (n = 51 subjects) of hyperpolarized 129Xe gas lung images transformed by common MRI artifacts (noise and nonlinear intensity distortion). The resulting ventilation‐based segmentations were used to assess algorithmic performance and characterize optimization domain differences in terms of measurement bias and precision. Results Although facilitating computational processing and providing discriminating clinically relevant measures of interest, histogram‐based segmentation methods discard important contextual spatial information and are consequently less robust in terms of measurement precision in the presence of common MRI artifacts relative to the image‐based convolutional neural network. Conclusions Direct optimization within the image domain using convolutional neural networks leverages spatial information, which mitigates problematic issues associated with histogram‐based approaches and suggests a preferred future research direction. Further, the entire processing and evaluation framework, including the newly reported deep learning functionality, is available as open source through the well‐known Advanced Normalization Tools ecosystem.

ANTsX neuroimaging-derived structural phenotypes of UK Biobank

“…Intensity-based data augmentation consisted of randomly added noise based on ITK functionality, simulated bias fields based on N4 bias field modeling [Tustison et al, 2010], and histogram warping for mimicking well-known MRI intensity nonlinearities [Nyúl and Udupa, 1999; Tustison et al, 2021a]. These augmentation techniques are available in ANTsXNet (Python and R):…”

“…Intensity-based data augmentation consisted of randomly added noise based on ITK functionality, simulated bias fields based on N4 bias field modeling [Tustison et al, 2010], and histogram warping for mimicking well-known MRI intensity nonlinearities [Nyúl and Udupa, 1999;Tustison et al, 2021a]. These augmentation techniques are available in ANTsXNet Additionally, given the well-studied variation of hippocampal volume with age [Bussy et al, 2021;Driscoll et al, 2003] and disease (e.g., [Henneman et al, 2009]), the original 17 image sets were warped to additional age/gender templates created from the public IXI database (i.e., 20-30, 30-40, 50-60, 70-80, 80-90 and male/female) and ADNI templates made from diagnostic cohorts (i.e., normal, mild cognitive impairment, and Alzheimer's disease).…”

A 3D stack‐of‐spirals approach for rapid hyperpolarized 129Xe ventilation mapping in pediatric cystic fibrosis lung disease

“…Phase‐encoding along k z was performed in a center‐out fashion to maximize the non‐renewable HP magnetization available near the center of k ‐space. Scan parameters, particularly those governing image resolution, were adjusted to be comparable to the 2D‐GRE acquisition (Table 1), while others were guided by previous spiral and non‐Cartesian 129 Xe MRI literature 12,17 . The acquisition time and consequent breath‐hold for the 3D‐SoS sequence ranged from 1.2–1.8 s, using the same number of prescribed slices as the 2D‐GRE scan.…”

3D Single-Breath Chemical Shift Imaging Hyperpolarized Xe-129 MRI of Healthy, CF, IPF, and COPD Subjects

“…Ventilation images were segmented into regions of no ventilation, hypoventilation, normal ventilation, and hyperventilation using Advanced Normalization Tools software package (ANTs) [ 27 , 28 ] and fused with proton images in ITK-SNAP [ 29 ]. We used a N4 bias field correction, a whole lung segmentation and a Gaussian mixture model with a Markov random field spatial prior modeling [ 30 ]. Regions of no ventilation and hypoventilation were grouped as ventilation defects (VD); regions of normal ventilation and hyperventilation were grouped as healthy/normal.…”