Virginia Newcombe

Efficient Multi-Scale 3D CNN with Fully Connected CRF for Accurate Brain Lesion Segmentation

HIGHLIGHTSAn efficient 11‐layers deep, multi‐scale, 3D CNN architecture.A novel training strategy that significantly boosts performance.The first employment of a 3D fully connected CRF for post‐processing.State‐of‐the‐art performance on three challenging lesion segmentation tasks.New insights into the automatically learned intermediate representations. ABSTRACT We propose a dual pathway, 11‐layers deep, three‐dimensional Convolutional Neural Network for the challenging task of brain lesion segmentation. The devised architecture is the result of an in‐depth analysis of the limitations of current networks proposed for similar applications. To overcome the computational burden of processing 3D medical scans, we have devised an efficient and effective dense training scheme which joins the processing of adjacent image patches into one pass through the network while automatically adapting to the inherent class imbalance present in the data. Further, we analyze the development of deeper, thus more discriminative 3D CNNs. In order to incorporate both local and larger contextual information, we employ a dual pathway architecture that processes the input images at multiple scales simultaneously. For post‐processing of the networks soft segmentation, we use a 3D fully connected Conditional Random Field which effectively removes false positives. Our pipeline is extensively evaluated on three challenging tasks of lesion segmentation in multi‐channel MRI patient data with traumatic brain injuries, brain tumours, and ischemic stroke. We improve on the state‐of‐the‐art for all three applications, with top ranking performance on the public benchmarks BRATS 2015 and ISLES 2015. Our method is computationally efficient, which allows its adoption in a variety of research and clinical settings. The source code of our implementation is made publicly available.

Use of T2-weighted magnetic resonance imaging of the optic nerve sheath to detect raised intracranial pressure

IntroductionThe dural sheath surrounding the optic nerve communicates with the subarachnoid space, and distends when intracranial pressure is elevated. Magnetic resonance imaging (MRI) is often performed in patients at risk for raised intracranial pressure (ICP) and can be used to measure precisely the diameter of optic nerve and its sheath. The objective of this study was to assess the relationship between optic nerve sheath diameter (ONSD), as measured using MRI, and ICP.MethodsWe conducted a retrospective blinded analysis of brain MRI images in a prospective cohort of 38 patients requiring ICP monitoring after severe traumatic brain injury (TBI), and in 36 healthy volunteers. ONSD was measured on T2-weighted turbo spin-echo fat-suppressed sequence obtained at 3 Tesla MRI. ICP was measured invasively during the MRI scan via a parenchymal sensor in the TBI patients.ResultsMeasurement of ONSD was possible in 95% of cases. The ONSD was significantly greater in TBI patients with raised ICP (>20 mmHg; 6.31 ± 0.50 mm, 19 measures) than in those with ICP of 20 mmHg or less (5.29 ± 0.48 mm, 26 measures; P < 0.0001) or in healthy volunteers (5.08 ± 0.52 mm; P < 0.0001). There was a significant relationship between ONSD and ICP (r = 0.71, P < 0.0001). Enlarged ONSD was a robust predictor of raised ICP (area under the receiver operating characteristic curve = 0.94), with a best cut-off of 5.82 mm, corresponding to a negative predictive value of 92%, and to a value of 100% when ONSD was less than 5.30 mm.ConclusionsWhen brain MRI is indicated, ONSD measurement on images obtained using routine sequences can provide a quantitative estimate of the likelihood of significant intracranial hypertension.

A Role for the Default Mode Network in the Bases of Disorders of Consciousness

Functional connectivity in the default mode network (DMN) is known to be reduced in patients with disorders of consciousness, to a different extent depending on their clinical severity. Nevertheless, the integrity of the structural architecture supporting this network and its relation with the exhibited functional disconnections are very poorly understood. We investigated the structural connectivity and white matter integrity of the DMN in patients with disorders of consciousness of varying clinical severity.

Analysis of acute traumatic axonal injury using diffusion tensor imaging

Traumatic axonal injury (TAI) contributes significantly to mortality and morbidity following traumatic brain injury (TBI), but is poorly characterized by conventional imaging techniques. Diffusion tensor imaging (DTI) may provide better detection as well as insights into the mechanisms of white matter injury. DTI data from 33 patients with moderate-to-severe TBI, acquired at a median of 32 h postinjury, were compared with data from 28 age-matched controls. The global burden of whole brain white matter injury (GBWMI) was quantified by measuring the proportion of voxels that lay below a critical fractional anisotropy (FA) threshold, identified from control data. Mechanisms of change in FA maps were explored using an Eigenvalue analysis of the diffusion tensor. When compared with controls, patients showed significantly reduced mean FA (p < 0.001) and increased apparent diffusion coefficient (ADC; p = 0.017). GBWMI was significantly greater in patients than in controls (p < 0.01), but did not distinguish patients with obvious white matter lesions seen on structural imaging. It predicted classification of DTI images as head injury with a high degree of accuracy. Eigenvalue analysis showed that reductions in FA were predominantly the result of increases in radial diffusivity (p < 0.001). DTI may help quantify the overall burden of white matter injury in TBI and provide insights into underlying pathophysiology. Eigenvalue analysis suggests that the early imaging changes seen in white matter are consistent with axonal swelling rather than axonal truncation. This technique holds promise for examining disease progression, and may help define therapeutic windows for the treatment of diffuse brain injury.

Aetiological differences in neuroanatomy of the vegetative state: insights from diffusion tensor imaging and functional implications

Background An improved in vivo understanding of variations in neuropathology in the vegetative state (VS) may aid diagnosis, improve prognostication and help refine the selection of patients for particular treatment regimes. The authors have used diffusion tensor imaging (DTI) to characterise the extent and location of white matter loss in VS secondary to traumatic brain injury (TBI) and ischaemic–hypoxic injury. Methods Twelve patients with VS (seven TBI, five ischaemic/hypoxic injuries) underwent MRI including DTI at a minimum of 3 months postinjury. Mean apparent diffusion coefficient, fractional anisotropy and eigenvalues were obtained for whole-brain grey and white matter, the pons, thalamus, ventral midbrain, dorsal midbrain and the corpus callosum. DTI measures of supratentorial damage were compared with a summed measure from the JFK modified Coma Recovery Scale (CRS-R) and with a three-point scale of functional magnetic resonance imaging (fMRI) response to an auditory paradigm to assess whether residual integrity of supratentorial white matter connectivity correlated with cortical processing. Results Conventional radiological approaches did not detect lesions in regions where quantitative DTI demonstrated abnormalities. There was evidence of marked, broadly similar, abnormalities in the supratentorial grey- and white-matter compartments from both aetiologies. In contrast, discordant findings were found in the infratentorial compartment, with DTI abnormalities in the brainstem confined to the TBI group. Supratentorial DTI abnormalities correlated with the CRS-R as well as responses to an fMRI paradigm that detected convert cognitive processing. Conclusions DTI may help to characterise differences in patients in VS. These findings may have implications for response to therapies, and should be taken into account in trials of interventions aimed at arousal in VS.

Parcellating the neuroanatomical basis of impaired decision-making in traumatic brain injury

Cognitive dysfunction is a devastating consequence of traumatic brain injury that affects the majority of those who survive with moderate-to-severe injury, and many patients with mild head injury. Disruption of key monoaminergic neurotransmitter systems, such as the dopaminergic system, may play a key role in the widespread cognitive dysfunction seen after traumatic axonal injury. Manifestations of injury to this system may include impaired decision-making and impulsivity. We used the Cambridge Gambling Task to characterize decision-making and risk-taking behaviour, outside of a learning context, in a cohort of 44 patients at least six months post-traumatic brain injury. These patients were found to have broadly intact processing of risk adjustment and probability judgement, and to bet similar amounts to controls. However, a patient preference for consistently early bets indicated a higher level of impulsiveness. These behavioural measures were compared with imaging findings on diffusion tensor magnetic resonance imaging. Performance in specific domains of the Cambridge Gambling Task correlated inversely and specifically with the severity of diffusion tensor imaging abnormalities in regions that have been implicated in these cognitive processes. Thus, impulsivity was associated with increased apparent diffusion coefficient bilaterally in the orbitofrontal gyrus, insula and caudate; abnormal risk adjustment with increased apparent diffusion coefficient in the right thalamus and dorsal striatum and left caudate; and impaired performance on rational choice with increased apparent diffusion coefficient in the bilateral dorsolateral prefrontal cortices, and the superior frontal gyri, right ventrolateral prefrontal cortex, the dorsal and ventral striatum, and left hippocampus. Importantly, performance in specific cognitive domains of the task did not correlate with diffusion tensor imaging abnormalities in areas not implicated in their performance. The ability to dissociate the location and extent of damage with performance on the various task components using diffusion tensor imaging allows important insights into the neuroanatomical basis of impulsivity following traumatic brain injury. The ability to detect such damage in vivo may have important implications for patient management, patient selection for trials, and to help understand complex neurocognitive pathways.

Mapping Traumatic Axonal Injury Using Diffusion Tensor Imaging: Correlations with Functional Outcome

Background Traumatic brain injury is a major cause of morbidity and mortality worldwide. Ameliorating the neurocognitive and physical deficits that accompany traumatic brain injury would be of substantial benefit, but the mechanisms that underlie them are poorly characterized. This study aimed to use diffusion tensor imaging to relate clinical outcome to the burden of white matter injury. Methodology/Principal Findings Sixty-eight patients, categorized by the Glasgow Outcome Score, underwent magnetic resonance imaging at a median of 11.8 months (range 6.6 months to 3.7 years) years post injury. Control data were obtained from 36 age-matched healthy volunteers. Mean fractional anisotropy, apparent diffusion coefficient (ADC), and eigenvalues were obtained for regions of interest commonly affected in traumatic brain injury. In a subset of patients where conventional magnetic resonance imaging was completely normal, diffusion tensor imaging was able to detect clear abnormalities. Significant trends of increasing ADC with worse outcome were noted in all regions of interest. In the white matter regions of interest worse clinical outcome corresponded with significant trends of decreasing fractional anisotropy. Conclusions/Significance This study found that clinical outcome was related to the burden of white matter injury, quantified by diffusivity parameters late after traumatic brain injury. These differences were seen even in patients with the best outcomes and patients in whom conventional magnetic resonance imaging was normal, suggesting that diffusion tensor imaging can detect subtle injury missed by other techniques. An improved in vivo understanding of the pathology of traumatic brain injury, including its distribution and extent, may enhance outcome evaluation and help to provide a mechanistic basis for deficits that remain unexplained by other approaches.

Unsupervised Domain Adaptation in Brain Lesion Segmentation with Adversarial Networks

Significant advances have been made towards building accurate automatic segmentation systems for a variety of biomedical applications using machine learning. However, the performance of these systems often degrades when they are applied on new data that differ from the training data, for example, due to variations in imaging protocols. Manually annotating new data for each test domain is not a feasible solution. In this work we investigate unsupervised domain adaptation using adversarial neural networks to train a segmentation method which is more robust to differences in the input data, and which does not require any annotations on the test domain. Specifically, we derive domain-invariant features by learning to counter an adversarial network, which attempts to classify the domain of the input data by observing the activations of the segmentation network. Furthermore, we propose a multi-connected domain discriminator for improved adversarial training. Our system is evaluated using two MR databases of subjects with traumatic brain injuries, acquired using different scanners and imaging protocols. Using our unsupervised approach, we obtain segmentation accuracies which are close to the upper bound of supervised domain adaptation.

Microstructural basis of contusion expansion in traumatic brain injury: insights from diffusion tensor imaging

Traumatic brain injury (TBI) is often exacerbated by events that lead to secondary brain injury, and represent potentially modifiable causes of mortality and morbidity. Diffusion tensor imaging was used to characterize tissue at-risk in a group of 35 patients scanned at a median of 50 hours after injury. Injury progression was assessed in a subset of 16 patients with two scans. All contusions within the first few days of injury showed a core of restricted diffusion, surrounded by an area of raised apparent diffusion coefficient (ADC). In addition to these two well-defined regions, a thinner rim of reduced ADC was observed surrounding the region of increased ADC in 91% of patients scanned within the first 3 days after injury. In patients who underwent serial imaging, the rim of ADC hypointensity was subsumed into the high ADC region as the contusion enlarged. Overall contusion enlargement tended to be more frequent with early lesions, but its extent was unrelated to the time of initial imaging, initial contusion size, or the presence of hemostatic abnormalities. This rim of hypointensity may characterize a region of microvascular failure resulting in cytotoxic edema, and may represent a ‘traumatic penumbra’ which may be rescued by effective therapy.

Pathophysiologic Mechanisms of Cerebral Ischemia and Diffusion Hypoxia in Traumatic Brain Injury.

IMPORTANCE Combined oxygen 15-labeled positron emission tomography (15O PET) and brain tissue oximetry have demonstrated increased oxygen diffusion gradients in hypoxic regions after traumatic brain injury (TBI). These data are consistent with microvascular ischemia and are supported by pathologic studies showing widespread microvascular collapse, perivascular edema, and microthrombosis associated with selective neuronal loss. Fluorine 18-labeled fluoromisonidazole ([18F]FMISO), a PET tracer that undergoes irreversible selective bioreduction within hypoxic cells, could confirm these findings. OBJECTIVE To combine [18F]FMISO and 15O PET to demonstrate the relative burden, distribution, and physiologic signatures of conventional macrovascular and microvascular ischemia in early TBI. DESIGN, SETTING, AND PARTICIPANTS This case-control study included 10 patients who underwent [18F]FMISO and 15O PET within 1 to 8 days of severe or moderate TBI. Two cohorts of 10 healthy volunteers underwent [18F]FMISO or 15O PET. The study was performed at the Wolfson Brain Imaging Centre of Addenbrookes Hospital. Cerebral blood flow, cerebral blood volume, cerebral oxygen metabolism (CMRO2), oxygen extraction fraction, and brain tissue oximetry were measured in patients during [18F]FMISO and 15O PET imaging. Similar data were obtained from control cohorts. Data were collected from November 23, 2007, to May 22, 2012, and analyzed from December 3, 2012, to January 6, 2016. MAIN OUTCOMES AND MEASURES Estimated ischemic brain volume (IBV) and hypoxic brain volume (HBV) and a comparison of their spatial distribution and physiologic signatures. RESULTS The 10 patients with TBI (9 men and 1 woman) had a median age of 59 (range, 30-68) years; the 2 control cohorts (8 men and 2 women each) had median ages of 53 (range, 41-76) and 45 (range, 29-59) years. Compared with controls, patients with TBI had a higher median IBV (56 [range, 9-281] vs 1 [range, 0-11] mL; P < .001) and a higher median HBV (29 [range, 0-106] vs 9 [range, 1-24] mL; P = .02). Although both pathophysiologic tissue classes were present within injured and normal appearing brains, their spatial distributions were poorly matched. When compared with tissue within the IBV compartment, the HBV compartment showed similar median cerebral blood flow (17 [range, 11-40] vs 14 [range, 6-22] mL/100 mL/min), cerebral blood volume (2.4 [range, 1.6- 4.2] vs 3.9 [range, 3.4-4.8] mL/100 mL), and CMRO2 (44 [range, 27-67] vs 71 [range, 34-88] μmol/100 mL/min) but a lower oxygen extraction fraction (38% [range, 29%-50%] vs 89% [range, 75%-100%]; P < .001), and more frequently showed CMRO2 values consistent with irreversible injury. Comparison with brain tissue oximetry monitoring suggested that the threshold for increased [18F]FMISO trapping is probably 15 mm Hg or lower. CONCLUSIONS AND RELEVANCE Tissue hypoxia after TBI is not confined to regions with structural abnormality and can occur in the absence of conventional macrovascular ischemia. This physiologic signature is consistent with microvascular ischemia and is a target for novel neuroprotective strategies.