Search

Brain functional connectivity (FC) reveals biomarkers for identification of various neuropsychiatric disorders. Recent application of deep neural networks (DNNs) to connectome-based classification mostly relies on traditional convolutional neural networks using input connectivity matrices on a regular Euclidean grid. We propose a graph deep learning framework to incorporate the non-Euclidean information about graph structure for classifying functional magnetic resonance imaging (fMRI)-derived brain networks in major depressive disorder (MDD). We design a novel graph autoencoder (GAE) architecture based on the graph convolutional networks (GCNs) to embed the topological structure and node content of large-sized fMRI networks into low-dimensional latent representations. In network construction, we employ the Ledoit-Wolf (LDW) shrinkage method to estimate the high-dimensional FC metrics efficiently from fMRI data. We consider both supervised and unsupervised approaches for the graph embedding learning. The learned embeddings are then used as feature inputs for a deep fully-connected neural network (FCNN) to discriminate MDD from healthy controls. Evaluated on two resting-state fMRI (rs-fMRI) MDD datasets, results show that the proposed GAE-FCNN model significantly outperforms several state-of-the-art methods for brain connectome classification, achieving the best accuracy using the LDW-FC edges as node features. The graph embeddings of fMRI FC networks learned by the GAE also reveal apparent group differences between MDD and HC. Our new framework demonstrates feasibility of learning graph embeddings on brain networks to provide discriminative information for diagnosis of brain disorders.

DesmondL Campbell,1 Hakmook Kang,2 Sepideh Shokouhi1 On behalf of The Alzheimer’s Disease Neuroimaging Initiative 1Department of Radiology and Radiological Sciences, 2Department of Biostatistics, Vanderbilt University Medical Center, Vanderbilt University Institute of Imaging Science, Nashville, TN, USA Objectives: Semi-quantitative image analysis methods in Alzheimer’s Disease (AD) require normalization of positron emission tomography (PET) images. However, recent studies have found variabilities associated with reference region selection of amyloid PET images. Haralick features (HFs) generated from the Gray Level Co-occurrence Matrix (GLCM) quantify spatial characteristics of amyloid PET radiotracer uptake without the need for intensity normalization. The objective of this study is to calculate several HFs in different diagnostic groups and determine the group differences.Methods: All image and metadata were acquired through the Alzheimer’s Disease Neuroimaging Initiative database. Subjects were grouped in three ways: by clinical diagnosis, by APOE e4 allele, and by Alzheimer’s Disease Assessment Scale-cognitive subscale (ADAS-Cog) score. Several GLCM matrices were calculated for different direction and distances (1–4 mm) from multiple regions on PET images. The HFs, contrast, correlation, dissimilarity, energy, entropy, and homogeneity, were calculated from these GLCMs. Wilcoxon tests and Student t-tests were performed on Haralick features and standardized uptake value ratio (SUVR) values, respectively, to determine group differences. In addition to statistical testing, receiver operating characteristic (ROC) curves were generated to determine the discrimination performance of the selected regional HFs and the SUVR values.Results: Preliminary results from statistical testing indicate that HFs were capable of distinguishing groups at baseline and follow-up (false discovery rate corrected p

Objective To assess associations between nonalcoholic fatty liver disease (NAFLD) and measures of brain health in a population-based sample of adults. Methods Participants from the CARDIA study (Y25 exam; age 43–55 years) with concurrent CT quantification of liver fat, visceral adipose tissue (VAT) and brain magnetic resonance (MR) images were included (n=505). NAFLD was identified after exclusion of other causes of liver fat. Total tissue volume (TTV) and gray matter cerebral blood flow (GM-CBF) were estimated using 3T brain MR images. Results NAFLD prevalence was 18%. NAFLD was associated with lower TTV and GM-CBF after adjusting for intracranial volume, demographics, and health behaviors (p

Purpose: Cell size is a fundamental characteristic of all tissues, and changes in cell size in cancer reflect tumor status and response to treatments, such as apoptosis and cell cycle arrest. Unfortunately, cell size can only be obtained by pathologic evaluation of the tumor in the current standard of care. Previous imaging approaches can be implemented on only animal MRI scanners or require relatively long acquisition times that are undesirable for clinical imaging. There is a need to develop cell size imaging for clinics. Experimental Design: We propose a new method, IMPULSED (Imaging Microstructural Parameters Using Limited Spectrally Edited Diffusion) that can characterize mean cell sizes in solid tumors. We report the use of combined sequences with different gradient waveforms on human MRI and analytical equations that link DWI signals of real gradient waveforms and specific microstructural parameters such as cell size. We also describe comprehensive validations using computer simulations, cell experiments in vitro, and animal experiments in vivo and finally demonstrate applications in pre-operative breast cancer patients. Results: With fast acquisitions (~ 7 mins), IMPULSED can provide high-resolution (1.3 mm in-plane) mapping of mean cell size of human tumors in vivo on currently-available 3T MRI scanners. All validations suggest IMPULSED provide accurate and reliable measurements of mean cell size. Conclusion: The proposed IMPULSED method can assess cell size variations in the tumor of breast cancer patients, which may have the potential to assess early response to neoadjuvant therapy.

BackgroundSemiquantitative methods such as the standardized uptake value ratio (SUVR) require normalization of the radiotracer activity to a reference tissue to monitor changes in the accumulation of amyloid-β (Aβ) plaques measured with positron emission tomography (PET). The objective of this study was to evaluate the effect of reference tissue normalization in a test-retest (18)F-florbetapir SUVR study using cerebellar gray matter, white matter (two different segmentation masks), brainstem, and corpus callosum as reference regions.MethodsWe calculated the correlation between (18)F-florbetapir PET and concurrent cerebrospinal fluid (CSF) Aβ1-42 levels in a late mild cognitive impairment cohort with longitudinal PET and CSF data over the course of 2 years. In addition to conventional SUVR analysis using mean and median values of normalized brain radiotracer activity, we investigated a new image analysis technique-the weighted two-point correlation function (wS2)-to capture potentially more subtle changes in Aβ-PET data.ResultsCompared with the SUVRs normalized to cerebellar gray matter, all cerebral-to-white matter normalization schemes resulted in a higher inverse correlation between PET and CSF Aβ1-42, while the brainstem normalization gave the best results (high and most stable correlation). Compared with the SUVR mean and median values, the wS2 values were associated with the lowest coefficient of variation and highest inverse correlation to CSF Aβ1-42 levels across all time points and reference regions, including the cerebellar gray matter.ConclusionsThe selection of reference tissue for normalization and the choice of image analysis method can affect changes in cortical (18)F-florbetapir uptake in longitudinal studies.

Previous studies have demonstrated how imaging of the breast with patients lying prone using a supportive positioning device markedly facilitates longitudinal and/or multimodal image registration. In this contribution, the authors' primary objective was to determine if there are differences in the standardized uptake value (SUV) derived from [(18)F]fluorodeoxyglucose (18F-FDG) positron emission tomography (PET) in breast tumors imaged in the standard supine position and in the prone position using a specialized positioning device.A custom positioning device was constructed to allow for breast scanning in the prone position. Rigid and nonrigid phantom studies evaluated differences in prone and supine PET. Clinical studies comprised 18F-FDG-PET of 34 patients with locally advanced breast cancer imaged in the prone position (with the custom support) followed by imaging in the supine position (without the support). Mean and maximum values (SUVpeak and SUVmax, respectively) were obtained from tumor regions-of-interest for both positions. Prone and supine SUV were linearly corrected to account for the differences in 18F-FDG uptake time. Correlation, Bland-Altman, and nonparametric analyses were performed on uptake time-corrected and uncorrected data.SUV from the rigid PET breast phantom imaged in the prone position with the support device was 1.9% lower than without the support device. In the nonrigid PET breast phantom, prone SUV with the support device was 5.0% lower than supine SUV without the support device. In patients, the median (range) difference in uptake time between prone and supine scans was 16.4 min (13.4-30.9 min), which was significantly-but not completely-reduced by the linear correction method. SUVpeak and SUVmax from prone versus supine scans were highly correlated, with concordance correlation coefficients of 0.91 and 0.90, respectively. Prone SUVpeak and SUVmax were significantly lower than supine in both original and uptake time-adjusted data across a range of index times (P0.0001, Wilcoxon signed rank test). Before correcting for uptake time differences, Bland-Altman analyses revealed proportional bias between prone and supine measurements (SUVpeak and SUVmax) that increased with higher levels of FDG uptake. After uptake time correction, this bias was significantly reduced (P0.01). Significant prone-supine differences, with regard to the spatial distribution of lesions relative to isocenter, were observed between the two scan positions, but this was poorly correlated with the residual (uptake time-corrected) prone-supine SUVpeak difference (P = 0.78).Quantitative 18F-FDG-PET/CT of the breast in the prone position is not deleteriously affected by the support device but yields SUV that is consistently lower than those obtained in the standard supine position. SUV differences between scans arising from FDG uptake time differences can be substantially reduced, but not removed entirely, with the current correction method. SUV from the two scan orientations is quantitatively different and should not be assumed equivalent or interchangeable within the same subject. These findings have clinical relevance in that they underscore the importance of patient positioning while scanning as a clinical variable that must be accounted for with longitudinal PET measurement, for example, in the assessment of treatment response.

The purpose of this study was to determine whether multiparametric magnetic resonance imaging (MRI) using dynamic contrast-enhanced MRI (DCE-MRI) and diffusion-weighted MRI (DWI), obtained before and after the first cycle of neoadjuvant chemotherapy (NAC), is superior to single-parameter measurements for predicting pathologic complete response (pCR) in patients with breast cancer.Patients with stage II/III breast cancer were enrolled in an institutional review board-approved study in which 3-T DCE-MRI and DWI data were acquired before (n = 42) and after 1 cycle (n = 36) of NAC. Estimates of the volume transfer rate (K), extravascular extracellular volume fraction (ve), blood plasma volume fraction (vp), and the efflux rate constant (kep = K/ve) were generated from the DCE-MRI data using the Extended Tofts-Kety model. The apparent diffusion coefficient (ADC) was estimated from the DWI data. The derived parameter kep/ADC was compared with single-parameter measurements for its ability to predict pCR after the first cycle of NAC.The kep/ADC after the first cycle of NAC discriminated patients who went on to achieve a pCR (P0.001) and achieved a sensitivity, specificity, positive predictive value, and area under the receiver operator curve (AUC) of 0.92, 0.78, 0.69, and 0.88, respectively. These values were superior to the single parameters kep (AUC, 0.76) and ADC (AUC, 0.82). The AUCs between kep/ADC and kep were significantly different on the basis of the bootstrapped 95% confidence intervals (0.018-0.23), whereas the AUCs between kep/ADC and ADC trended toward significance (-0.11 to 0.24).The multiparametric analysis of DCE-MRI and DWI was superior to the single-parameter measurements for predicting pCR after the first cycle of NAC.