Splenic Hypoplasia

Our objective is to express time-varying wFNCs as weighted sums of correlation patterns whose contributions change independently of each other in time (see Fig 3(A) and 3(B)), allowing us to develop a richer picture of the interplay between connectivity patterns that are strongly present in the data. This objective explicitly permits collections of basis connectivity patterns featuring:
To achieve this goal we apply group temporal independent component analysis (tICA) (Fig 3(A)) [30 (link)] to wFNC matrices concatenated along the subject×time dimension, decomposing this concatenated 1081 network-pair correlations × 136 time windows × 314 subject structure into five maximally mutually independent timecourses (because we are performing this analysis at the group level, these are in fact length 136∙314 = 42,704 subject×time “courses”), each with an associated 47×47 connectivity pattern (a modular component of the mixing matrix) that is shared across subjects (Fig 3(A)). In the text above, for convenience, we will refer to the connectivity patterns as components, even though it is the subject×timecourses that are being estimated by tICA. In this decomposition, individual wFNCs are specified as weighted sums of the five CPs, yielding a 5-dimensional characterization of each subject’s 1081-dimensional connectivity structure in each time window. The dynamical object of investigation is now a set of 136 time-indexed five-vectors per subject that representing the contributions of five 1081-dimensional CPs to the observed wFNCs. The tICA decomposition of wFNC data produces CPs whose weights in each time-indexed five-vector are maximally mutually independent [28 (link)–30 (link)]. These tICA CPs are thus patterns whose additive contributions to observed wFNCs “pipe in” and fade out in a relatively independent manner. Although the window-wise CP weights in the tICA decomposition are as independent as possible, intrinsic dependencies within the data ensure that the weights are not formally independent, i.e. $P(a_1\le w_1\le b_1,\dots ,a_5\le w_5\le b_5)\ne \prod _{k=1}^{5}P(a_k\le w_k\le b_k)$ . The five-vectors thus hold information not available by analyzing elements separately, but maximizing temporal independence keeps the state-space from collapsing onto a lower-dimensional space, ie. if CP #1 and CP #2 are systematically mutually dependent then only one is necessary and the state space becomes four-dimensional. Reducing the systematic dependencies between CPs ensures we are taking maximal advantage of the dimensionality in which the dynamics have been defined.
Although we have chosen to focus on maximally temporally independent correlation patterns produced by applying temporal ICA to the windowed FNCs (Fig 3(A)), we were interested in understanding how sensitive the results obtained might be to our choice of method for extracting correlation patterns from the windowed FNCs. Thus we performed the same analysis on correlation patterns obtained from three other commonly utilized data-driven methods: spatial independent component analysis (sICA), principal component analysis (PCA) and kmeans clustering. We also explored the role of model order within the temporal ICA framework by repeating our analysis for temporal ICAs producing 2, 3, 6 and 7 (small perturbations of the featured 5 correlation pattern case).

An area under the receiver operating characteristic (AUROC) test was used to assess the internal and external validity and the chunk- and branch-level accuracy. First, we tested the performance of the algorithm in a set of independent stroke patients with ICAS. The structural dissimilarities in the cerebral arterial configuration between the healthy standard controls and those with stroke with ICAS were expected to indicate the clinical relevance of this algorithm in subjects with pathological conditions.

All of the patients underwent a carotid ultrasound (US) exam, including a Contrast-Enhanced Ultrasound (CEUS) study, within 15 days from the occurrence of the index stroke. An experienced examiner, P.C.-R., certified in Neurosonology by the Spanish Society of Neurology, performed the US examinations using a Philips CX50^® Ultrasound Machine (Philips, Amsterdam, Netherlands) with a linear probe. The US study protocol consisted of two parts: a standard B-mode and color-Doppler carotid plaque characterization, and the CEUS examination.
The extracranial common carotid arteries and the ICAs were examined in the longitudinal and the transverse planes. A plaque was defined as a localized lumen narrowing of ≥1.5 mm or an increase of >50% in the intima-media thickness compared to the adjacent portion of the vessel wall. When a plaque was identified, the following sonographic variables were recorded: (1) morphology of the plaque (concentric or eccentric); (2) echogenicity of the plaque classified as I = uniformly hypoechoic, II = predominantly hypoechoic, III = predominantly hyperechoic, IV = uniformly hyperechoic; and V = calcified plaque [32 (link)]; and (3) degree of stenosis by hemodynamic criteria [27 (link)]. Only the largest plaque was studied when patients presented more than one plaque in the ICA. For the statistical analysis, the echogenicity of the plaque was also classified as predominantly hypoechoic (<50% of the surface, comprising I and II categories) and predominantly hyperechoic (≥50% of the surface, comprising III, IV, and V categories).
After the B-mode and color-Doppler characterization, the CEUS examination was performed using the preset real-time, contrast-enhanced imaging modality with coded pulse inversion from the Philips CX50^® Ultrasound Machine. This setting decreases the mechanical index to 0.1, obtaining an almost completely black screen in the absence of contrast. Then, a bolus of 2 mL of Sonovue^® contrast was injected into a peripheral vein and flushed with 10 mL of saline according to the recommendations of the manufacturer (Bracco Imaging, Milan, Italy). At that point, the lumen was filled with the hyperechoic bubbles of the contrast defining the perimeter of the plaque in negative. Time gain compensation was adjusted to achieve homogeneous signal intensity. Finally, a DICOM cine loop was recorded for 120 s starting when the contrast bolus was injected and plaque neovessels were identified as hyperechoic bubbles appearing within the plaque perimeter.
The neovascularization of plaque was classified into grades: 0 (no visible microbubbles within the plaque), 1 (moderate microbubbles confined to the shoulder and/or adventitial side of the plaque), and 2 (diffuse microbubbles throughout the plaque), as previously described [33 (link)]. This grading was performed by P.C.-R. and by F.C., who was blinded to all of the clinical information, to calculate the interrater agreement and the Cohen’s kappa coefficient. When a discrepancy was detected, the images were reviewed, and a consensus between raters was required. A plaque was suitable for analysis of neovascularization if a DICOM cine loop of 15 s presented enough quality without movement artifact (for example swallowing) and at least 50% of the plaque was visible without calcium shadows.