Extracting morphologies from third harmonic generation images of structurally normal human brain tissue.

MOTIVATION: The morphologies contained in 3D third harmonic generation (THG) images of human brain tissue can report on the pathological state of the tissue. However, the complexity of THG brain images makes the usage of modern image processing tools, especially those of image filtering, segmentation and validation, to extract this information challenging. RESULTS: We developed a salient edge-enhancing model of anisotropic diffusion for image filtering, based on higher order statistics. We split the intrinsic 3-phase segmentation problem into two 2-phase segmentation problems, each of which we solved with a dedicated model, active contour weighted by prior extreme. We applied the novel proposed algorithms to THG images of structurally normal ex-vivo human brain tissue, revealing key tissue components-brain cells, microvessels and neuropil, enabling statistical characterization of these components. Comprehensive comparison to manually delineated ground truth validated the proposed algorithms. Quantitative comparison to second harmonic generation/auto-fluorescence images, acquired simultaneously from the same tissue area, confirmed the correctness of the main THG features detected. AVAILABILITY AND IMPLEMENTATION: The software and test datasets are available from the authors. CONTACT: z.zhang@vu.nl. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

Disentangling Somatosensory Evoked Potentials of the Fingers: Limitations and Clinical Potential.

In searching for clinical biomarkers of the somatosensory function, we studied reproducibility of somatosensory potentials (SEP) evoked by finger stimulation in healthy subjects. SEPs induced by electrical stimulation and especially after median nerve stimulation is a method widely used in the literature. It is unclear, however, if the EEG recordings after finger stimulation are reproducible within the same subject. We tested in five healthy subjects the consistency and reproducibility of responses through bootstrapping as well as test-retest recordings. We further evaluated the possibility to discriminate activity of different fingers both at electrode and at source level. The lack of consistency and reproducibility suggest responses to finger stimulation to be unreliable, even with reasonably high signal-to-noise ratio and adequate number of trials. At sources level, somatotopic arrangement of the fingers representation was only found in one of the subjects. Although finding distinct locations of the different fingers activation was possible, our protocol did not allow for non-overlapping dipole representations of the fingers. We conclude that despite its theoretical advantages, we cannot recommend the use of somatosensory potentials evoked by finger stimulation to extract clinical biomarkers.

Determination of head conductivity frequency response in vivo with optimized EIT-EEG.

Electroencephalography (EEG) benefits from accurate head models. Dipole source modelling errors can be reduced from over 1cm to a few millimetres by replacing generic head geometry and conductivity with tailored ones. When adequate head geometry is available, electrical impedance tomography (EIT) can be used to infer the conductivities of head tissues. In this study, the boundary element method (BEM) is applied with three-compartment (scalp, skull and brain) subject-specific head models. The optimal injection of small currents to the head with a modular EIT current injector, and voltage measurement by an EEG amplifier is first sought by simulations. The measurement with a 64-electrode EEG layout is studied with respect to three noise sources affecting EIT: background EEG, deviations from the fitting assumption of equal scalp and brain conductivities, and smooth model geometry deviations from the true head geometry. The noise source effects were investigated depending on the positioning of the injection and extraction electrode and the number of their combinations used sequentially. The deviation from equal scalp and brain conductivities produces rather deterministic errors in the three conductivities irrespective of the current injection locations. With a realistic measurement of around 2 min and around 8 distant distinct current injection pairs, the error from the other noise sources is reduced to around 10% or less in the skull conductivity. The analysis of subsequent real measurements, however, suggests that there could be subject-specific local thinnings in the skull, which could amplify the conductivity fitting errors. With proper analysis of multiplexed sinusoidal EIT current injections, the measurements on average yielded conductivities of 340 mS/m (scalp and brain) and 6.6 mS/m (skull) at 2 Hz. From 11 to 127 Hz, the conductivities increased by 1.6% (scalp and brain) and 6.7% (skull) on the average. The proper analysis was ensured by using recombination of the current injections into virtual ones, avoiding problems in location-specific skull morphology variations. The observed large intersubject variations support the need for in vivo measurement of skull conductivity, resulting in calibrated subject-specific head models.

Alzheimer Disease and Behavioral Variant Frontotemporal Dementia: Automatic Classification Based on Cortical Atrophy for Single-Subject Diagnosis.

Purpose To investigate the diagnostic accuracy of an image-based classifier to distinguish between Alzheimer disease (AD) and behavioral variant frontotemporal dementia (bvFTD) in individual patients by using gray matter (GM) density maps computed from standard T1-weighted structural images obtained with multiple imagers and with independent training and prediction data. Materials and Methods The local institutional review board approved the study. Eighty-four patients with AD, 51 patients with bvFTD, and 94 control subjects were divided into independent training (n = 115) and prediction (n = 114) sets with identical diagnosis and imager type distributions. Training of a support vector machine (SVM) classifier used diagnostic status and GM density maps and produced voxelwise discrimination maps. Discriminant function analysis was used to estimate suitability of the extracted weights for single-subject classification in the prediction set. Receiver operating characteristic (ROC) curves and area under the ROC curve (AUC) were calculated for image-based classifiers and neuropsychological z scores. Results Training accuracy of the SVM was 85% for patients with AD versus control subjects, 72% for patients with bvFTD versus control subjects, and 79% for patients with AD versus patients with bvFTD (P ≤ .029). Single-subject diagnosis in the prediction set when using the discrimination maps yielded accuracies of 88% for patients with AD versus control subjects, 85% for patients with bvFTD versus control subjects, and 82% for patients with AD versus patients with bvFTD, with a good to excellent AUC (range, 0.81-0.95; P ≤ .001). Machine learning-based categorization of AD versus bvFTD based on GM density maps outperforms classification based on neuropsychological test results. Conclusion The SVM can be used in single-subject discrimination and can help the clinician arrive at a diagnosis. The SVM can be used to distinguish disease-specific GM patterns in patients with AD and those with bvFTD as compared with normal aging by using common T1-weighted structural MR imaging. (©) RSNA, 2015.

A three domain covariance framework for EEG/MEG data.

In this paper we introduce a covariance framework for the analysis of single subject EEG and MEG data that takes into account observed temporal stationarity on small time scales and trial-to-trial variations. We formulate a model for the covariance matrix, which is a Kronecker product of three components that correspond to space, time and epochs/trials, and consider maximum likelihood estimation of the unknown parameter values. An iterative algorithm that finds approximations of the maximum likelihood estimates is proposed. Our covariance model is applicable in a variety of cases where spontaneous EEG or MEG acts as source of noise and realistic noise covariance estimates are needed, such as in evoked activity studies, or where the properties of spontaneous EEG or MEG are themselves the topic of interest, like in combined EEG-fMRI experiments in which the correlation between EEG and fMRI signals is investigated. We use a simulation study to assess the performance of the estimator and investigate the influence of different assumptions about the covariance factors on the estimated covariance matrix and on its components. We apply our method to real EEG and MEG data sets.

Are Epilepsy-Related fMRI Components Dependent on the Presence of Interictal Epileptic Discharges in Scalp EEG?

Spatial independent component analysis (ICA) is increasingly being used to extract resting-state networks from fMRI data. Previous studies showed that ICA also reveals independent components (ICs) related to the seizure onset zone. However, it is currently unknown how these epileptic ICs depend on the presence of interictal epileptic discharges (IEDs) in the EEG. The goal of this study was to explore the relation between ICs obtained from fMRI epochs during the occurrence of IEDs in the EEG and those without IEDs. fMRI data sets with co-registered EEG were retrospectively selected of patients from whom the location of the epileptogenic zone was confirmed by outcome of surgery (n = 8). The fMRI data were split into two epochs: one with IEDs visible in scalp EEG and one without. Spatial ICA was applied to the fMRI data of each part separately. The maps of all resulting components were compared to the resection area and the EEG-fMRI correlation pattern by computing a spatial correlation coefficient to detect the epilepsy-related component. For all patients, except one, there was a remarkable resemblance between the epilepsy-related components selected during epochs with IEDs and those without IEDs. These findings suggest that epilepsy-related ICs are not dependent on the presence of IEDs in scalp EEG. Since these epileptic ICs showed partial overlap with resting-state networks of healthy volunteers (n = 10), our study supports the need for new ways to classify epileptic ICs.

Brain network alterations in Alzheimer's disease measured by eigenvector centrality in fMRI are related to cognition and CSF biomarkers.

Recent imaging studies have demonstrated functional brain network changes in patients with Alzheimer's disease (AD). Eigenvector centrality (EC) is a graph analytical measure that identifies prominent regions in the brain network hierarchy and detects localized differences between patient populations. This study used voxel-wise EC mapping (ECM) to analyze individual whole-brain resting-state functional magnetic resonance imaging (MRI) scans in 39 AD patients (age 67 ± 8) and 43 healthy controls (age 69 ± 7). Between-group differences were assessed by a permutation-based method. Associations of EC with biomarkers for AD pathology in cerebrospinal fluid (CSF) and Mini Mental State Examination (MMSE) scores were assessed using Spearman correlation analysis. Decreased EC was found bilaterally in the occipital cortex in AD patients compared to controls. Regions of increased EC were identified in the anterior cingulate and paracingulate gyrus. Across groups, frontal and occipital EC changes were associated with pathological concentrations of CSF biomarkers and with cognition. In controls, decreased EC values in the occipital regions were related to lower MMSE scores. Our main finding is that ECM, a hypothesis-free and computationally efficient analysis method of functional MRI (fMRI) data, identifies changes in brain network organization in AD patients that are related to cognition and underlying AD pathology. The relation between AD-like EC changes and cognitive performance suggests that resting-state fMRI measured EC is a potential marker of disease severity for AD.

Novel artefact removal algorithms for co-registered EEG/fMRI based on selective averaging and subtraction.

Co-registered EEG and functional MRI (EEG/fMRI) is a potential clinical tool for planning invasive EEG in patients with epilepsy. In addition, the analysis of EEG/fMRI data provides a fundamental insight into the precise physiological meaning of both fMRI and EEG data. Routine application of EEG/fMRI for localization of epileptic sources is hampered by large artefacts in the EEG, caused by switching of scanner gradients and heartbeat effects. Residuals of the ballistocardiogram (BCG) artefacts are similarly shaped as epileptic spikes, and may therefore cause false identification of spikes. In this study, new ideas and methods are presented to remove gradient artefacts and to reduce BCG artefacts of different shapes that mutually overlap in time. Gradient artefacts can be removed efficiently by subtracting an average artefact template when the EEG sampling frequency and EEG low-pass filtering are sufficient in relation to MR gradient switching (Gonçalves et al., 2007). When this is not the case, the gradient artefacts repeat themselves at time intervals that depend on the remainder between the fMRI repetition time and the closest multiple of the EEG acquisition time. These repetitions are deterministic, but difficult to predict due to the limited precision by which these timings are known. Therefore, we propose to estimate gradient artefact repetitions using a clustering algorithm, combined with selective averaging. Clustering of the gradient artefacts yields cleaner EEG for data recorded during scanning of a 3T scanner when using a sampling frequency of 2048 Hz. It even gives clean EEG when the EEG is sampled with only 256 Hz. Current BCG artefacts-reduction algorithms based on average template subtraction have the intrinsic limitation that they fail to deal properly with artefacts that overlap in time. To eliminate this constraint, the precise timings of artefact overlaps were modelled and represented in a sparse matrix. Next, the artefacts were disentangled with a least squares procedure. The relevance of this approach is illustrated by determining the BCG artefacts in a data set consisting of 29 healthy subjects recorded in a 1.5 T scanner and 15 patients with epilepsy recorded in a 3 T scanner. Analysis of the relationship between artefact amplitude, duration and heartbeat interval shows that in 22% (1.5T data) to 30% (3T data) of the cases BCG artefacts show an overlap. The BCG artefacts of the EEG/fMRI data recorded on the 1.5T scanner show a small negative correlation between HBI and BCG amplitude. In conclusion, the proposed methodology provides a substantial improvement of the quality of the EEG signal without excessive computer power or additional hardware than standard EEG-compatible equipment.

EEG-fMRI correlation patterns in the presurgical evaluation of focal epilepsy: a comparison with electrocorticographic data and surgical outcome measures.

EEG-correlated functional MRI (EEG-fMRI) visualizes brain regions associated with interictal epileptiform discharges (IEDs). This technique images the epileptiform network, including multifocal, superficial and deeply situated cortical areas. To understand the role of EEG-fMRI in presurgical evaluation, its results should be validated relative to a gold standard. For that purpose, EEG-fMRI data were acquired for a heterogeneous group of surgical candidates (n=16) who were later implanted with subdural grids and strips (ECoG). The EEG-fMRI correlation patterns were systematically compared with brain areas involved in IEDs ECoG, using a semi-automatic analysis method, as well as to the seizure onset zone, resected area, and degree of seizure freedom. In each patient at least one of the EEG-fMRI areas was concordant with an interictally active ECoG area, always including the early onset area of IEDs in the ECoG data. This confirms that EEG-fMRI reflects a pattern of onset and propagation of epileptic activity. At group level, 76% of the BOLD regions that were covered with subdural grids, were concordant with interictally active ECoG electrodes. Due to limited spatial sampling, 51% of the BOLD regions were not covered with electrodes and could, therefore, not be validated. From an ECoG perspective it appeared that 29% of the interictally active ECoG regions were missed by EEG-fMRI and that 68% of the brain regions were correctly identified as inactive with EEG-fMRI. Furthermore, EEG-fMRI areas included the complete seizure onset zone in 83% and resected area in 93% of the data sets. No clear distinction was found between patients with a good or poor surgical outcome: in both patient groups, EEG-fMRI correlation patterns were found that were either focal or widespread. In conclusion, by comparison of EEG-fMRI with interictal invasive EEG over a relatively large patient population we were able to show that the EEG-fMRI correlation patterns are spatially accurate at the level of neurosurgical units (i.e. anatomical brain regions) and reflect the underlying network of IEDs. Therefore, we expect that EEG-fMRI can play an important role for the determination of the implantation strategy.

A semi-automatic method to determine electrode positions and labels from gel artifacts in EEG/fMRI-studies.

The analysis of simultaneous EEG and fMRI data is generally based on the extraction of regressors of interest from the EEG, which are correlated to the fMRI data in a general linear model setting. In more advanced approaches, the spatial information of EEG is also exploited by assuming underlying dipole models. In this study, we present a semi automatic and efficient method to determine electrode positions from electrode gel artifacts, facilitating the integration of EEG and fMRI in future EEG/fMRI data models. In order to visualize all electrode artifacts simultaneously in a single view, a surface rendering of the structural MRI is made using a skin triangular mesh model as reference surface, which is expanded to a "pancake view". Then the electrodes are determined with a simple mouse click for each electrode. Using the geometry of the skin surface and its transformation to the pancake view, the 3D coordinates of the electrodes are reconstructed in the MRI coordinate frame. The electrode labels are attached to the electrode positions by fitting a template grid of the electrode cap in which the labels are known. The correspondence problem between template and sample electrodes is solved by minimizing a cost function over rotations, shifts and scalings of the template grid. The crucial step here is to use the solution of the so-called "Hungarian algorithm" as a cost function, which makes it possible to identify the electrode artifacts in arbitrary order. The template electrode grid has to be constructed only once for each cap configuration. In our implementation of this method, the whole procedure can be performed within 15 min including import of MRI, surface reconstruction and transformation, electrode identification and fitting to template. The method is robust in the sense that an electrode template created for one subject can be used without identification errors for another subject for whom the same EEG cap was used. Furthermore, the method appears to be robust against spurious or missing artifacts. We therefore consider the proposed method as a useful and reliable tool within the larger toolbox required for the analysis of co-registered EEG/fMRI data.

Sensory handedness is not reflected in cortical responses after basic nerve stimulation: a MEG study.

Motor dominance is well established, but sensory dominance is much less clear. We therefore studied the cortical evoked magnetic fields using magnetoencephalography (MEG) in a group of 20 healthy right handed subjects in order to examine whether standard electrical stimulation of the median and ulnar nerve demonstrated sensory lateralization. The global field power (GFP) curves, as an indication of cortical activation, did not depict sensory lateralization to the dominant left hemisphere. Comparison of the M20, M30, and M70 peak latencies and GFP values exhibited no statistical differences between the hemispheres, indicating no sensory hemispherical dominance at these latencies for each nerve. Field maps at these latencies presented a first and second polarity reversal for both median and ulnar stimulation. Spatial dipole position parameters did not reveal statistical left-right differences at the M20, M30 and M70 peaks for both nerves. Neither did the dipolar strengths at M20, M30 and M70 show a statistical left-right difference for both nerves. Finally, the Laterality Indices of the M20, M30 and M70 strengths did not indicate complete lateralization to one of the hemispheres. After electrical median and ulnar nerve stimulation no evidence was found for sensory hand dominance in brain responses of either hand, as measured by MEG. The results can provide a new assessment of patients with sensory dysfunctions or perceptual distortion when sensory dominance occurs way beyond the estimated norm.

On the cutting edge of glioblastoma surgery: where neurosurgeons agree and disagree on surgical decisions.

OBJECTIVE: The aim of glioblastoma surgery is to maximize the extent of resection while preserving functional integrity. Standards are lacking for surgical decision-making, and previous studies indicate treatment variations. These shortcomings reflect the need to evaluate larger populations from different care teams. In this study, the authors used probability maps to quantify and compare surgical decision-making throughout the brain by 12 neurosurgical teams for patients with glioblastoma. METHODS: The study included all adult patients who underwent first-time glioblastoma surgery in 2012-2013 and were treated by 1 of the 12 participating neurosurgical teams. Voxel-wise probability maps of tumor location, biopsy, and resection were constructed for each team to identify and compare patient treatment variations. Brain regions with different biopsy and resection results between teams were identified and analyzed for patient functional outcome and survival. RESULTS: The study cohort consisted of 1087 patients, of whom 363 underwent a biopsy and 724 a resection. Biopsy and resection decisions were generally comparable between teams, providing benchmarks for probability maps of resections and biopsies for glioblastoma. Differences in biopsy rates were identified for the right superior frontal gyrus and indicated variation in biopsy decisions. Differences in resection rates were identified for the left superior parietal lobule, indicating variations in resection decisions. CONCLUSIONS: Probability maps of glioblastoma surgery enabled capture of clinical practice decisions and indicated that teams generally agreed on which region to biopsy or to resect. However, treatment variations reflecting clinical dilemmas were observed and pinpointed by using the probability maps, which could therefore be useful for quality-of-care discussions between surgical teams for patients with glioblastoma.

Data-driven modeling of phase interactions between spontaneous MEG oscillations.

OBJECTIVE: Synchronization between distributed rhythms in the brain is commonly assessed by estimating the synchronization strength from simultaneous measurements. This approach, however, does not elucidate the phase dynamics that underlies synchronization. For this, an explicit dynamical model is required. Based on the assumption that the recorded rhythms can be described as weakly coupled oscillators, we propose a method for characterizing their phase-interaction dynamics. METHODS: We propose to model ongoing magnetoencephalographic (MEG) oscillations as weakly coupled oscillators. Based on this model, the phase interactions between simultaneously recorded signals are characterized by estimating the modulation in instantaneous frequency as a function of their phase difference. Furthermore, we mathematically derive the effect of volume conduction on the model and show how indices for strength and direction of coupling can be derived. RESULTS: The methodology is tested using simulations and is applied to ongoing occipital-frontal MEG oscillations of healthy subjects in the alpha and beta bands during rest. The simulations show that the model is robust against the presence of noise, short observation times, and model violations. The application to MEG data shows that the model can reconstruct the observed occipital-frontal phase difference distributions. Furthermore, it suggests that phase locking in the alpha and beta band is established by qualitatively different mechanisms. CONCLUSION: When the recorded rhythms are assumed to be weakly coupled oscillators, a dynamical model for the phase interactions can be fitted to data. The model is able to reconstruct the observed phase difference distribution, and hence, provides a dynamical explanation for observed phase locking.

Anesthetic block of pain-related cortical activity in patients with peripheral nerve injury measured by magnetoencephalography.

BACKGROUND: This study examined whether chronic neuropathic pain, modulated by a local anesthetic block, is associated with cortical magnetic field changes. METHODS: In a group of 20 patients with pain caused by unilateral traumatic peripheral nerve injury, a local block with lidocaine 1% was administered and the cortical effects were measured and compared with a control group. The global field power (GFP), describing distribution of cortical activation after median and ulnar nerve stimulation, was plotted and calculated. The effects on the affected hemisphere and the unaffected hemisphere (UH) before and after a block of the injured nerve were statistically evaluated. RESULTS: Major differences based on the GFP curves, at a component between 50 ms - 90 ms (M70), were found in patients: in the affected hemisphere the M70 GFP peak values were statistically significantly larger in comparison with the UH, and the GFP curves differed morphologically. Interestingly, the mean UH responses were reduced in comparison with the control group, a finding suggesting that the UH is also part of the cortical changes. At M70, the GFP curves and values in the affected hemisphere were modulated by a local block of the median or the ulnar nerve. The most likely location of cortical adaptation is in the primary somatosensory cortex. CONCLUSIONS: Cortical activation is enhanced in the affected hemisphere compared with the UH and is modulated by a local block. The UH in neuropathic pain changes as well. Evoked fields may offer an opportunity to monitor the effectiveness of treatments of neuropathic pain in humans.

Correction for pulse height variability reduces physiological noise in functional MRI when studying spontaneous brain activity.

EEG correlated functional MRI (EEG-fMRI) allows the delineation of the areas corresponding to spontaneous brain activity, such as epileptiform spikes or alpha rhythm. A major problem of fMRI analysis in general is that spurious correlations may occur because fMRI signals are not only correlated with the phenomena of interest, but also with physiological processes, like cardiac and respiratory functions. The aim of this study was to reduce the number of falsely detected activated areas by taking the variation in physiological functioning into account in the general linear model (GLM). We used the photoplethysmogram (PPG), since this signal is based on a linear combination of oxy- and deoxyhemoglobin in the arterial blood, which is also the basis of fMRI. We derived a regressor from the variation in pulse height (VIPH) of PPG and added this regressor to the GLM. When this regressor was used as predictor it appeared that VIPH explained a large part of the variance of fMRI signals acquired from five epilepsy patients and thirteen healthy volunteers. As a confounder VIPH reduced the number of activated voxels by 30% for the healthy volunteers, when studying the generators of the alpha rhythm. Although for the patients the number of activated voxels either decreased or increased, the identification of the epileptogenic zone was substantially enhanced in one out of five patients, whereas for the other patients the effects were smaller. In conclusion, applying VIPH as a confounder diminishes physiological noise and allows a more reliable interpretation of fMRI results.

Accurate MR Image Registration to Anatomical Reference Space for Diffuse Glioma.

To summarize the distribution of glioma location within a patient population, registration of individual MR images to anatomical reference space is required. In this study, we quantified the accuracy of MR image registration to anatomical reference space with linear and non-linear transformations using estimated tumor targets of glioblastoma and lower-grade glioma, and anatomical landmarks at pre- and post-operative time-points using six commonly used registration packages (FSL, SPM5, DARTEL, ANTs, Elastix, and NiftyReg). Routine clinical pre- and post-operative, post-contrast T1-weighted images of 20 patients with glioblastoma and 20 with lower-grade glioma were collected. The 2009a Montreal Neurological Institute brain template was used as anatomical reference space. Tumors were manually segmented in the patient space and corresponding healthy tissue was delineated as a target volume in the anatomical reference space. Accuracy of the tumor alignment was quantified using the Dice score and the Hausdorff distance. To measure the accuracy of general brain alignment, anatomical landmarks were placed in patient and in anatomical reference space, and the landmark distance after registration was quantified. Lower-grade gliomas were registered more accurately than glioblastoma. Registration accuracy for pre- and post-operative MR images did not differ. SPM5 and DARTEL registered tumors most accurate, and FSL least accurate. Non-linear transformations resulted in more accurate general brain alignment than linear transformations, but tumor alignment was similar between linear and non-linear transformation. We conclude that linear transformation suffices to summarize glioma locations in anatomical reference space.

Robust Deep Learning-based Segmentation of Glioblastoma on Routine Clinical MRI Scans Using Sparsified Training.

PURPOSE: To improve the robustness of deep learning-based glioblastoma segmentation in a clinical setting with sparsified datasets. MATERIALS AND METHODS: In this retrospective study, preoperative T1-weighted, T2-weighted, T2-weighted fluid-attenuated inversion recovery, and postcontrast T1-weighted MRI from 117 patients (median age, 64 years; interquartile range [IQR], 55-73 years; 76 men) included within the Multimodal Brain Tumor Image Segmentation (BraTS) dataset plus a clinical dataset (2012-2013) with similar imaging modalities of 634 patients (median age, 59 years; IQR, 49-69 years; 382 men) with glioblastoma from six hospitals were used. Expert tumor delineations on the postcontrast images were available, but for various clinical datasets, one or more sequences were missing. The convolutional neural network, DeepMedic, was trained on combinations of complete and incomplete data with and without site-specific data. Sparsified training was introduced, which randomly simulated missing sequences during training. The effects of sparsified training and center-specific training were tested using Wilcoxon signed rank tests for paired measurements. RESULTS: A model trained exclusively on BraTS data reached a median Dice score of 0.81 for segmentation on BraTS test data but only 0.49 on the clinical data. Sparsified training improved performance (adjusted P < .05), even when excluding test data with missing sequences, to median Dice score of 0.67. Inclusion of site-specific data during sparsified training led to higher model performance Dice scores greater than 0.8, on par with a model based on all complete and incomplete data. For the model using BraTS and clinical training data, inclusion of site-specific data or sparsified training was of no consequence. CONCLUSION: Accurate and automatic segmentation of glioblastoma on clinical scans is feasible using a model based on large, heterogeneous, and partially incomplete datasets. Sparsified training may boost the performance of a smaller model based on public and site-specific data.Supplemental material is available for this article.Published under a CC BY 4.0 license.

Quantifying eloquent locations for glioblastoma surgery using resection probability maps.

OBJECTIVE: Decisions in glioblastoma surgery are often guided by presumed eloquence of the tumor location. The authors introduce the "expected residual tumor volume" (eRV) and the "expected resectability index" (eRI) based on previous decisions aggregated in resection probability maps. The diagnostic accuracy of eRV and eRI to predict biopsy decisions, resectability, functional outcome, and survival was determined. METHODS: Consecutive patients with first-time glioblastoma surgery in 2012-2013 were included from 12 hospitals. The eRV was calculated from the preoperative MR images of each patient using a resection probability map, and the eRI was derived from the tumor volume. As reference, Sawaya's tumor location eloquence grades (EGs) were classified. Resectability was measured as observed extent of resection (EOR) and residual volume, and functional outcome as change in Karnofsky Performance Scale score. Receiver operating characteristic curves and multivariable logistic regression were applied. RESULTS: Of 915 patients, 674 (74%) underwent a resection with a median EOR of 97%, functional improvement in 71 (8%), functional decline in 78 (9%), and median survival of 12.8 months. The eRI and eRV identified biopsies and EORs of at least 80%, 90%, or 98% better than EG. The eRV and eRI predicted observed residual volumes under 10, 5, and 1 ml better than EG. The eRV, eRI, and EG had low diagnostic accuracy for functional outcome changes. Higher eRV and lower eRI were strongly associated with shorter survival, independent of known prognostic factors. CONCLUSIONS: The eRV and eRI predict biopsy decisions, resectability, and survival better than eloquence grading and may be useful preoperative indices to support surgical decisions.

Tensor regularized total variation for denoising of third harmonic generation images of brain tumors.

Third harmonic generation (THG) microscopy shows great potential for instant pathology of brain tissue during surgery. However, the rich morphologies contained and the noise associated makes image restoration, necessary for quantification of the THG images, challenging. Anisotropic diffusion filtering (ADF) has been recently applied to restore THG images of normal brain, but ADF is hard-to-code, time-consuming and only reconstructs salient edges. This work overcomes these drawbacks by expressing ADF as a tensor regularized total variation model, which uses the Huber penalty and the L1 norm for tensor regularization and fidelity measurement, respectively. The diffusion tensor is constructed from the structure tensor of ADF yet the tensor decomposition is performed only in the non-flat areas. The resulting model is solved by an efficient and easy-to-code primal-dual algorithm. Tests on THG brain tumor images show that the proposed model has comparable denoising performance as ADF while it much better restores weak edges and it is up to 60% more time efficient.

FAst Segmentation Through SURface Fairing (FASTSURF): A novel semi-automatic hippocampus segmentation method.

OBJECTIVE: The objective is to present a proof-of-concept of a semi-automatic method to reduce hippocampus segmentation time on magnetic resonance images (MRI). MATERIALS AND METHODS: FAst Segmentation Through SURface Fairing (FASTSURF) is based on a surface fairing technique which reconstructs the hippocampus from sparse delineations. To validate FASTSURF, simulations were performed in which sparse delineations extracted from full manual segmentations served as input. On three different datasets with different diagnostic groups, FASTSURF hippocampi were compared to the original segmentations using Jaccard overlap indices and percentage volume differences (PVD). In one data set for which back-to-back scans were available, unbiased estimates of overlap and PVD were obtained. Using longitudinal scans, we compared hippocampal atrophy rates measured by manual, FASTSURF and two automatic segmentations (FreeSurfer and FSL-FIRST). RESULTS: With only seven input contours, FASTSURF yielded mean Jaccard indices ranging from 72(±4.3)% to 83(±2.6)% and PVDs ranging from 0.02(±2.40)% to 3.2(±3.40)% across the three datasets. Slightly poorer results were obtained for the unbiased analysis, but the performance was still considerably better than both tested automatic methods with only five contours. CONCLUSIONS: FASTSURF segmentations have high accuracy and require only a fraction of the delineation effort of fully manual segmentation. Atrophy rate quantification based on completely manual segmentation is well reproduced by FASTSURF. Therefore, FASTSURF is a promising tool to be implemented in clinical workflow, provided a future prospective validation confirms our findings.