No abstract

Purpose: Diffusion-weighted imaging allows investigators to identify structural, microstructural, and connectivity-based differences between subjects, but variability due to session and scanner biases is a challenge. Methods: To investigate DWI variability, we present MASiVar, a multisite data set consisting of 319 diffusion scans acquired at 3 T from b = 1000 to 3000 s/mm 2 across 14 healthy adults, 83 healthy children (5 to 8 years), three sites, and four scanners as a publicly available, preprocessed, and de-identified data set. With the adult data, we demonstrate the capacity of MASiVar to simultaneously quantify the intrasession, intersession, interscanner, and intersubject variability of four common DWI processing approaches: (1) a tensor signal representation, (2) a multi-compartment neurite orientation dispersion and density model, (3) white-matter bundle segmentation, and (4) structural connectomics. Respectively, we evaluate region-wise fractional anisotropy, mean diffusivity, and principal eigenvector; region-wise CSF volume fraction, intracellular volume fraction, and orientation dispersion index; bundle-wise shape, | 3305 CAI et Al. F I G U R E 8Overall trends in coefficient of variation (CoV) across DTI, NODDI, bundle segmentation, and connectomics. Visualization of median CoV across all four processing approaches on the intrasession, intersession, interscanner, and intersubject levels illustrates consistently increased variability with session, scanner, and subject effects. Statistical significance was determined with the Wilcoxon signed-rank test with and without Bonferroni correction. The outlying points correspond to the NODDI cVF approach in white matter where absolute cVF values are expected to be low.

No abstract

Diffusion weighted imaging (DWI) allows investigators to identify microstructural differences between subjects, but variability due to session and scanner biases is still a challenge. To investigate DWI variability, we present MASiVar, a multisite dataset consisting of 319 diffusion scans acquired at 3T from b = 1000 to 3000 s/mm2 across 97 different healthy subjects and four different scanners as a publicly available, preprocessed, and de-identified dataset. With these data we characterize variability on the intrasession intrascanner (N = 158), intersession intrascanner (N = 328), intersession interscanner (N = 53), and intersubject intrascanner (N = 80) levels. Our baseline analysis focuses on four common DWI processing approaches: (1) a tensor signal representation, (2) a multi-compartment neurite orientation dispersion and density model, (3) white matter bundle segmentation, and (4) structural connectomics. Respectively, we evaluate region-wise fractional anisotropy (FA), mean diffusivity, and principal eigenvector; region-wise cerebral spinal fluid volume fraction, intracellular volume fraction, and orientation dispersion index; bundle-wise shape, volume, length and FA; and connectome correlation and maximized modularity, global efficiency, and characteristic path length. We plot the scan/re-scan discrepancies in these measures at each level and find that variability generally increases with intrasession to intersession to interscanner to intersubject effects and that sometimes interscanner variability can approach intersubject variability. This baseline study suggests harmonization between scanners for multisite analyses is critical prior to inference of group differences on subjects and demonstrates the potential of MASiVar to investigate DWI variability across multiple levels and processing approaches simultaneously.

Multiple instance learning (MIL) is a supervised learning methodology that aims to allow models to learn instance class labels from bag class labels, where a bag is defined to contain multiple instances. MIL is gaining traction for learning from weak labels but has not been widely applied to 3D medical imaging. MIL is well-suited to clinical CT acquisitions since (1) the highly anisotropic voxels hinder application of traditional 3D networks and (2) patch-based networks have limited ability to learn whole volume labels. In this work, we apply MIL with a deep convolutional neural network to identify whether clinical CT head image volumes possess one or more large hemorrhages (> 20cm 3 ), resulting in a learned 2D model without the need for 2D slice annotations. Individual image volumes are considered separate bags, and the slices in each volume are instances. Such a framework sets the stage for incorporating information obtained in clinical reports to help train a 2D segmentation approach. Within this context, we evaluate the data requirements to enable generalization of MIL by varying the amount of training data. Our results show that a training size of at least 400 patient image volumes was needed to achieve accurate per-slice hemorrhage detection. Over a five-fold cross-validation, the leading model, which made use of the maximum number of training volumes, had an average true positive rate of 98.10%, an average true negative rate of 99.36%, and an average precision of 0.9698. The models have been made available along with source code 1 to enabled continued exploration and adaption of MIL in CT neuroimaging.