Validated imaging biomarkers as decision-making tools in clinical trials and routine practice: current status and recommendations from the EIBALL* subcommittee of the European Society of Radiology (ESR)

Observer-driven pattern recognition is the standard for interpretation of medical images. To achieve global parity in interpretation, semi-quantitative scoring systems have been developed based on observer assessments; these are widely used in scoring coronary artery disease, the arthritides and neurological conditions and for indicating the likelihood of malignancy. However, in an era of machine learning and artificial intelligence, it is increasingly desirable that we extract quantitative biomarkers from medical images that inform on disease detection, characterisation, monitoring and assessment of response to treatment. Quantitation has the potential to provide objective decision-support tools in the management pathway of patients. Despite this, the quantitative potential of imaging remains under-exploited because of variability of the measurement, lack of harmonised systems for data acquisition and analysis, and crucially, a paucity of evidence on how such quantitation potentially affects clinical decision-making and patient outcome. This article reviews the current evidence for the use of semi-quantitative and quantitative biomarkers in clinical settings at various stages of the disease pathway including diagnosis, staging and prognosis, as well as predicting and detecting treatment response. It critically appraises current practice and sets out recommendations for using imaging objectively to drive patient management decisions.

MRI language dominance assessment in epilepsy patients at 1.0 T: region of interest analysis and comparison with intracarotid amytal testing

The primary goal of this study was to test the reliability of presurgical language lateralization in epilepsy patients with functional magnetic resonance imaging (fMRI) with a 1.0-T MR scanner using a simple word generation paradigm and conventional equipment. In addition, hemispherical fMRI language lateralization analysis and region of interest (ROI) analysis in the frontal and temporo-parietal regions were compared with the intracarotid amytal test (IAT). Twenty epilepsy patients under presurgical evaluation were prospectively examined by both fMRI and IAT. The fMRI experiment consisted of a word chain task (WCT) using the conventional headphone set and a sparse sequence. In 17 of the 20 patients, data were available for comparison between the two procedures. Fifteen of these 17 patients were categorized as left hemispheric dominant, and 2 patients demonstrated bilateral language representation by both fMRI and IAT. The highest reliability for lateralization was obtained using frontal ROI analysis. Hemispherical analysis was less powerful and reliable in all cases but one, while temporo-parietal ROI analysis was unreliable as a stand-alone analysis when compared with IAT. The effect of statistical threshold on language lateralization prompted for the use of t-value-dependent lateralization index plots. This study illustrates that fMRI-determined language lateralization can be performed reliably in a clinical MR setting operating at a low field strength of 1 T without expensive stimulus presentation systems.

31P-NMR spectroscopy and the metabolic properties of different muscle fibers

To study the in vivo recruitment of different fiber types and their metabolic properties, 31P-nuclear magnetic resonance spectroscopy (31P-NMRS) of the human calf muscle was performed in seven normal sedentary subjects. In the exhaustive exercise protocol used, the work load was increased every minute during 5 min. This resulted in a prominent split of the Pi resonance in all subjects, indicating pH compartmentation in the muscles studied. From the chemical shift of the Pi peaks relative to phosphocreatine (PCr) at the end of the exercise, intracellular pH (pHi) averaged 6.92 +/- 0.05 (SD) in compartment 1 and 6.23 +/- 0.15 in compartment 2. The recovery of both Pi resonances after exercise could be followed easily in five of these subjects. The recovery rate of the Pi peak is a good estimate of the oxidative metabolism at the end of the exercise. A monoexponential regression analysis showed that the mean initial recovery rate S0 was 2.49 +/- 0.17%/s in compartment 1 and only 0.87 +/- 0.12%/s in compartment 2, indicating aerobic function three times higher in compartment 1 at the end of exercise. The mean relative ATP fraction dropped significantly (P less than 0.001), from 20.0 +/- 1.0% of the total 31P signal integral before exercise to 14.0 +/- 1.6% at the end of exercise. The simultaneous visualization of two compartments, in good order, one with high pHi and fast recovery and another with low pHi and slow recovery, is rationalized by the different metabolic behavior of type I and II fibers in human calf muscle in response to exhaustive exercise. This study demonstrates that 31P-NMRS is an excellent noninvasive procedure to quantify aerobic metabolism in both fiber types simultaneously.

Artefacts in multi-echo T2 imaging for high-precision gel dosimetry: II. Analysis of B1-field inhomogeneity

In BANG gel dosimetry, the spin-spin relaxation rate, R2 = 1/T2, is related to radiation dose that has been delivered to a gel phantom. R2 is calculated by fitting the pixel intensities of a set of differently T2-weighted base images. The accuracy that is aimed for in this quantitative MR application is about 5% relative to the maximum dose. In a conventional imaging MR scanner, however, several imaging artefacts may perturb the final dose map. These deviations manifest themselves as either a deformation of the dose map or an inaccuracy of the dose pixel value. Inaccuracies in the dose maps are caused by both spatial and temporal deviations in signal intensities during scanning. This study deals with B1-field inhomogeneities as a source of dose inaccuracy. First, the influence of B1-field inhomogeneities on slice profiles is investigated using a thin-slice phantom. Secondly, a FLASH sequence is used to map the B1-field by assessing the effective flip angle in each voxel of a homogeneous phantom. In addition, both experiments and computer simulations revealed the effects of B1 field inhomogeneities on the measured R2. This work offers a method to correct R2 maps for B1 -field inhomogeneities.

Magnetic resonance spectroscopy in migraine: what have we learned so far?

OBJECTIVE

To summarize and evaluate proton ((1)H) and phosphorus ((31)P) magnetic resonance spectroscopy (MRS) findings in migraine.

METHODS

A thorough review of (1)H and/or (31)P-MRS studies in any form of migraine published up to September 2011.

RESULTS

Some findings were consistent in all studies, such as a lack of ictal/interictal brain pH change and a disturbed energy metabolism, the latter of which is reflected in a drop in phosphocreatine content, both in the resting brain and in muscle following exercise. In a recent interictal study ATP was found to be significantly decreased in the occipital lobe of migraine with aura patients, reinforcing the concept of a mitochondrial component to the migraine threshold, at least in a subgroup of patients. In several studies a correlation between the extent of the energy disturbance and the clinical phenotype severity was apparent. Less consistent but still congruent with a disturbed energy metabolism is an observed lactate increase in the occipital cortex of several migraine subtypes (MwA, migraine with prolonged aura). No increases in brain glutamate levels were found.

CONCLUSION

The combined abnormalities found in MRS studies imply a mitochondrial component in migraine neurobiology. This could be due to a primary mitochondrial dysfunction or be secondary to, for example, alterations in brain excitability. The extent of variation in the data can be attributed to both the variable clinical inclusion criteria used and the variation in applied methodology. Therefore it is necessary to continue to optimize MRS methodology to gain further insights, especially concerning lactate and glutamate.

Muscle carnosine loading by beta-alanine supplementation is more pronounced in trained vs. untrained muscles

Carnosine occurs in high concentrations in human skeletal muscle and assists working capacity during high-intensity exercise. Chronic beta-alanine (BA) supplementation has consistently been shown to augment muscle carnosine concentration, but the effect of training on the carnosine loading efficiency is poorly understood. The aim of the present study was to compare muscle carnosine loading between trained and untrained arm and leg muscles. In a first study ( n = 17), reliability of carnosine quantification by proton magnetic resonance spectroscopy (1H-MRS) was evaluated in deltoid and triceps brachii muscles. In a second study, participants ( n = 35; 10 nonathletes, 10 cyclists, 10 swimmers, and 5 kayakers) were supplemented with 6.4 g/day of slow-release BA for 23 days. Carnosine content was evaluated in soleus, gastrocnemius medialis, and deltoid muscles by 1H-MRS. All the results are reported as arbitrary units. In the nonathletes, BA supplementation increased carnosine content by 47% in the arm and 33% in the leg muscles (not significant). In kayakers, the increase was more pronounced in arm (deltoid) vs. leg (soleus + gastrocnemius) muscles (0.089 vs. 0.049), whereas the reverse pattern was observed in cyclists (0.065 vs. 0.084). Swimmers had significantly higher increase in carnosine in both deltoid (0.107 vs. 0.065) and gastrocnemius muscle (0.082 vs. 0.051) compared with nonathletes. We showed that 1) carnosine content can be reliably measured by 1H-MRS in deltoid muscle, 2) carnosine loading is equally effective in arm vs. leg muscles of nonathletes, and 3) carnosine loading is more pronounced in trained vs. untrained muscles.

Consensus statement on current and emerging methods for the diagnosis and evaluation of cerebrovascular disease

Cerebrovascular disease (CVD) remains a leading cause of death and the leading cause of adult disability in most developed countries. This work summarizes state-of-the-art, and possible future, diagnostic and evaluation approaches in multiple stages of CVD, including (i) visualization of sub-clinical disease processes, (ii) acute stroke theranostics, and (iii) characterization of post-stroke recovery mechanisms. Underlying pathophysiology as it relates to large vessel steno-occlusive disease and the impact of this macrovascular disease on tissue-level viability, hemodynamics (cerebral blood flow, cerebral blood volume, and mean transit time), and metabolism (cerebral metabolic rate of oxygen consumption and pH) are also discussed in the context of emerging neuroimaging protocols with sensitivity to these factors. The overall purpose is to highlight advancements in stroke care and diagnostics and to provide a general overview of emerging research topics that have potential for reducing morbidity in multiple areas of CVD.

Dysfunctional modulation of default mode network activity in attention-deficit/hyperactivity disorder

The state regulation deficit model posits that individuals with attention-deficit/hyperactivity disorder (ADHD) have difficulty applying mental effort effectively under suboptimal conditions such as very fast and very slow event rates (ERs). ADHD is also associated with diminished suppression of default mode network (DMN) activity and related performance deficits on tasks requiring effortful engagement. The current study builds on these 2 literatures to test the hypothesis that failure to modulate DMN activity in ADHD might be especially pronounced at ER extremes. Nineteen adults with ADHD and 20 individuals without any neuropsychiatric condition successfully completed a simple target detection task under 3 ER conditions (2-, 4-, and 8-s interstimulus intervals) inside the scanner. Task-related DMN deactivations were compared between 2 groups. There was a differential effect of ER on DMN activity for individuals with ADHD compared to controls. Individuals with ADHD displayed excessive DMN activity at the fast and slow, but not at the moderate ER. The results indicate that DMN attenuation in ADHD is disrupted in suboptimal energetic states where additional effort is required to optimize task engagement. DMN dysregulation may be an important element of the neurobiological underpinnings of state regulation deficits in ADHD.

Artefacts in multi-echo T2 imaging for high-precision gel dosimetry: I. Analysis and compensation of eddy currents

In BANG gel dosimetry, the spin-spin relaxation rate, R2 = I/T2, is related to the radiation dose that has been delivered to the gel phantom. R2 is calculated by fitting the pixel intensities of a set of differently T2-weighted base images. In gel dosimetry for radiotherapy, an accuracy of 5% in dose and 3 mm spatially, whichever is lower, is the objective. Therefore, possible sources of artefacts must be considered and dealt with. To obtain a set of base images a multiple spin-echo sequence is used. However, in a conventional MR scanner eddy currents will be provoked by switching the imaging gradients. As the eddy currents change in the course of the sequence, the net magnetization will be affected accordingly. Hence, eddy currents may have a significant influence on the quantitative R2 images themselves as well as on their slice position. In this study, we report an analysis of the eddy currents as they appear in the multiple spin-echo sequence. Eddy currents are measured using a frequency shift method resulting in eddy current field maps. The related geometrical displacements are obtained by use of a pyramidal phantom. The R2 versus dose relation is determined in the three main directions of the magnet, revealing a dependence of the measured R2 on slice orientation. The time course of eddy currents is then used in a computer simulation to estimate the effects they produce in the recorded R2 images. A compensation method for eddy current effects in multi-echo T2 mapping is described.

Neural correlates of pantomiming familiar and unfamiliar tools: action semantics versus mechanical problem solving?

This study aims to reveal the neural correlates of planning and executing tool use pantomimes and explores the brain's response to pantomiming the use of unfamiliar tools. Sixteen right-handed volunteers planned and executed pantomimes of equally graspable familiar and unfamiliar tools while undergoing fMRI. During the planning of these pantomimes, we found bilateral temporo-occipital and predominantly left hemispheric frontal and parietal activation. The execution of the pantomimes produced additional activation in frontal and sensorimotor regions. In the left posterior parietal region both familiar and unfamiliar tool pantomimes elicit peak activity in the anterior portion of the lateral bank of the intraparietal sulcus--A region associated with the representation of action goals. The cerebral activation during these pantomimes is remarkably similar for familiar and unfamiliar tools, and direct comparisons revealed only few differences. First, the left cuneus is significantly active during the planning of pantomimes of unfamiliar tools, reflecting increased visual processing of the novel objects. Second, executing (but not planning) familiar tool pantomimes showed significant activation on the convex portion of the inferior parietal lobule, a region believed to serve as a repository for skilled object-related gestures. Given the striking similarity in brain activation while pantomiming familiar and unfamiliar tools, we argue that normal subjects use both action semantics and function from structure inferences simultaneously and interactively to give rise to flexible object-to-goal directed behavior.

Ictal source localization in presurgical patients with refractory epilepsy

Source localization of epileptic foci using ictal spatiotemporal dipole modeling (ISDM) yields reliable anatomic information in presurgical candidates. It requires substantial resources from EEG and neuroimaging laboratories. The profile and number of patients who may benefit from it are currently unknown. The purpose of this study is to demonstrate the clinical usefulness of source localization in a prospectively analyzed series. One hundred patients (51 male and 49 female patients) with mean age of 31 years (range, 2 to 63 years) and mean duration of refractory epilepsy of 20 years (range, 1 to 49 years) were enrolled consecutively in a presurgical protocol. Ictal EEG was available in 93 patients. ISDM was performed when suitable ictal EEG files were available. The clinical applicability of ISDM was examined in three patients groups: 37 patients in whom ictal EEG recording and MRI were congruent (group I), 30 patients in whom results were not completely congruent but not incongruent (group II), and 26 patients in whom the results were incongruent (group III). ISDM could be performed in 31 of 100 patients: 11 in group I, 8 in group II, and 12 in group III. ISDM influenced decision making in none of the patients in group I, in 4 of 8 patients in group II, and in 10 of 12 patients in group III. Typically, the results of ISDM directed avoiding intracranial EEG recordings in what appeared to be unsuitable candidates for resection by clearly confirming the incongruency between ictal EEG and MRI findings. In this series of 100 presurgical candidates, ictal source localization could be performed in 31% of patients. In 14% of patients, it proved to be a key element in the surgical decision process.

Presurgical evaluation of the motor hand area with functional MR imaging in patients with tumors and dysplastic lesions

PURPOSE

To test an optimized functional magnetic resonance (MR) imaging procedure to depict the motor hand representation area (HRA) in patients with epilepsy lesions near the central sulcus.

MATERIALS AND METHODS

Fast low-angle shot MR imaging was performed with an oblique single-section imaging technique in eight control subjects (10 hemispheres) and six patients (12 hemispheres). Three series of five activation images (obtained while subjects performed repetitive finger-to-thumb opposition movements) and five rest images were acquired. Each hemisphere was studied in three adjacent sections. Difference maps (obtained with simple subtraction between activation and rest images) were compared with t-test maps.

RESULTS

In control subjects, the HRA was visible in 27 of 30 sections. Qualitatively, activation was seen better on t-test maps in 14 and on difference maps in four of these sections. In all patients, motor activation could be seen in the hemisphere that contained the lesion. This activation was considered normal in four patients. In two patients, the HRA was deformed. Functional MR imaging activation in the motor area was confirmed with Penfield stimulation in five patients.

CONCLUSION

Functional MR imaging findings in the preoperative assessment of dysplastic lesions around the central sulcus are the same as for tumors. t-test maps are superior to difference maps in the treatment of motor functional MR imaging data.

The Vulnerability to Suicidal Behavior is Associated with Reduced Connectivity Strength

Suicidal behavior constitutes a major public health problem. Based on the stress–diathesis model, biological correlates of a diathesis might help to predict risk after stressor-exposure. Structural changes in cortical and subcortical areas and their connections have increasingly been linked with the diathesis. The current study identified structural network changes associated with a diathesis using a whole-brain approach by examining the structural connectivity between regions in euthymic suicide attempters (SA). In addition, the association between connectivity measures, clinical and genetic characteristics was investigated. We hypothesized that SA showed lower connectivity strength, associated with an increased severity of general clinical characteristics and an elevated expression of short alleles in serotonin polymorphisms. Thirteen euthymic SA were compared with fifteen euthymic non-attempters and seventeen healthy controls (HC). Clinical characteristics and three serotonin-related genetic polymorphisms were assessed. Diffusion MRI together with anatomical scans were administered. Preprocessing was performed using Explore DTI. Whole brain tractography of the diffusion-weighted images was followed by a number of streamlines-weighted network analysis using NBS. The network analysis revealed decreased connectivity strength in SA in the connections between the left olfactory cortex and left anterior cingulate gyrus. Furthermore, SA had increased suicidal ideation, hopelessness and self-reported depression, but did not show any differences for the genetic polymorphisms. Finally, lower connectivity strength between the right calcarine fissure and the left middle occipital gyrus was associated with increased trait anxiety severity (r_s = −0.78, p < 0.01) and hopelessness (r_s = −0.76, p < 0.01). SA showed differences in white matter network connectivity strength associated with clinical characteristics. Together, these variables could play an important role in predicting suicidal behavior.

Comparison of unsupervised classification methods for brain tumor segmentation using multi-parametric MRI

Tumor segmentation is a particularly challenging task in high-grade gliomas (HGGs), as they are among the most heterogeneous tumors in oncology. An accurate delineation of the lesion and its main subcomponents contributes to optimal treatment planning, prognosis and follow-up. Conventional MRI (cMRI) is the imaging modality of choice for manual segmentation, and is also considered in the vast majority of automated segmentation studies. Advanced MRI modalities such as perfusion-weighted imaging (PWI), diffusion-weighted imaging (DWI) and magnetic resonance spectroscopic imaging (MRSI) have already shown their added value in tumor tissue characterization, hence there have been recent suggestions of combining different MRI modalities into a multi-parametric MRI (MP-MRI) approach for brain tumor segmentation. In this paper, we compare the performance of several unsupervised classification methods for HGG segmentation based on MP-MRI data including cMRI, DWI, MRSI and PWI. Two independent MP-MRI datasets with a different acquisition protocol were available from different hospitals. We demonstrate that a hierarchical non-negative matrix factorization variant which was previously introduced for MP-MRI tumor segmentation gives the best performance in terms of mean Dice-scores for the pathologic tissue classes on both datasets.

•

Unsupervised classification algorithms are applied for brain tumor segmentation on multi-parametric MRI datasets.

•

Reported mean Dice-scores are in the range of state-of-the-art segmentation algorithms.

•

Hierarchical NMF obtained the best segmentation results in terms of mean Dice-scores for most of the tissue classes.

Small animal positron emission tomography during vagus nerve stimulation in rats: a pilot study

Vagus nerve stimulation (VNS) is an effective neurophysiological treatment for patients with refractory epilepsy, however, the mechanism of action remains unclear. Small animal positron emission tomography (PET) permits the monitoring of biochemical processes during multiple scans in the same animal. The aim of this pilot study was to explore the potential of 2-[18F]-fluoro-2-deoxy-d-glucose (FDG)-PET to investigate the effect of acute and chronic VNS on glucose metabolism in the rat brain. One week after EEG and VNS electrode implantation, a baseline FDG-PET scan was acquired during which animals were not stimulated. Secondly, scans were taken after first activation of the VNS electrode (acute VNS) and after one week of continuous VNS (chronic VNS). On the same time points, images were obtained in a control group. After acquisition, PET images were manually fused with MRI data. Normalized brain activities and left/right activity ratios of different brain structures were compared between control measurements and VNS group. During acute VNS, glucose metabolism was significantly decreased in the left hippocampus (P<0.05). Significant increases were found in both olfactory bulbs (P<0.05). During chronic VNS, a significant decrease in left/right ratio in the striatum (P<0.05) was found. Acute and chronic VNS induced changes in glucose metabolism in regions important for seizure control (hippocampus and striatum). Our results promote further brain research on VNS using small animal PET in rats.

Lyme disease in a child presenting with bilateral facial nerve palsy: MRI findings and review of the literature

We report a 7-year-old boy with neuroborreliosis presenting with headache and bilateral facial nerve palsy. MRI demonstrated tentorial and bilateral facial and trigeminal nerve enhancement.

Does visual cortex lactate increase following photic stimulation in migraine without aura patients? A functional (1)H-MRS study

Proton magnetic resonance spectroscopy ((1)H-MRS) has been used in a number of studies to assess noninvasively the temporal changes of lactate (Lac) in the activated human brain. Migraine neurobiology involves lack of cortical habituation to repetitive stimuli and a mitochondrial component has been put forward. Our group has recently demonstrated a reduction in the high-energy phosphates adenosine triphosphate (ATP) and phosphocreatine (PCr) in the occipital lobe of migraine without aura (MwoA) patients, at least in a subgroup, in a phosphorus MRS ((31)P-MRS) study. In previous studies, basal Lac levels or photic stimulation (PS)-induced Lac levels were found to be increased in patients with migraine with aura (MwA) and migraine patients with visual symptoms and paraesthesia, paresia and/or dysphasia, respectively. The aim of this study was to perform functional (1)H-MRS at 3 T in 20 MwoA patients and 20 control subjects. Repetitive visual stimulation was applied using MR-compatible goggles with 8 Hz checkerboard stimulation during 12 min. We did not observe any significant differences in signal integrals, ratios and absolute metabolite concentrations, including Lac, between MwoA patients and controls before PS. Lac also did not increase significantly during and following PS, both for MwoA patients and controls. Subtle Lac changes, smaller than the sensitivity threshold (i.e. estimated at 0.1-0.2 μmol/g at 3 T), cannot be detected by MRS. Our study does, however, argue against a significant switch to non-aerobic glucose metabolism during long-lasting PS of the visual cortex in MwoA patients.

"Unforgettable" - a pictorial essay on anatomy and pathology of the hippocampus

Abstract

The hippocampus is a small but complex anatomical structure that plays an important role in spatial and episodic memory. The hippocampus can be affected by a wide range of congenital variants and degenerative, inflammatory, vascular, tumoral and toxic-metabolic pathologies. Magnetic resonance imaging is the preferred imaging technique for evaluating the hippocampus. The main indications requiring tailored imaging sequences of the hippocampus are medically refractory epilepsy and dementia. The purpose of this pictorial review is threefold: (1) to review the normal anatomy of the hippocampus on MRI; (2) to discuss the optimal imaging strategy for the evaluation of the hippocampus; and (3) to present a pictorial overview of the most common anatomic variants and pathologic conditions affecting the hippocampus.

Teaching points

• Knowledge of normal hippocampal anatomy helps recognize anatomic variants and hippocampal pathology.

• Refractory epilepsy and dementia are the main indications requiring dedicated hippocampal imaging.

• Pathologic conditions centered in and around the hippocampus often have similar imaging features.

• Clinical information is often necessary to come to a correct diagnosis or an apt differential.

Anticipatory processes in brain state switching - evidence from a novel cued-switching task implicating default mode and salience networks

The default mode network (DMN) is the core brain system supporting internally oriented cognition. The ability to attenuate the DMN when switching to externally oriented processing is a prerequisite for effective performance and adaptive self-regulation. Right anterior insula (rAI), a core hub of the salience network (SN), has been proposed to control the switching from DMN to task-relevant brain networks. Little is currently known about the extent of anticipatory processes subserved by DMN and SN during switching. We investigated anticipatory DMN and SN modulation using a novel cued-switching task of between-state (rest-to-task/task-to-rest) and within-state (task-to-task) transitions. Twenty healthy adults performed the task implemented in an event-related functional magnetic resonance imaging (fMRI) design. Increases in activity were observed in the DMN regions in response to cues signalling upcoming rest. DMN attenuation was observed for rest-to-task switch cues. Obversely, DMN was up-regulated by task-to-rest cues. The strongest rAI response was observed to rest-to-task switch cues. Task-to-task switch cues elicited smaller rAI activation, whereas no significant rAI activation occurred for task-to-rest switches. Our data provide the first evidence that DMN modulation occurs rapidly and can be elicited by short duration cues signalling rest- and task-related state switches. The role of rAI appears to be limited to certain switch types - those implicating transition from a resting state and to tasks involving active cognitive engagement.

Semi-automated brain tumor segmentation on multi-parametric MRI using regularized non-negative matrix factorization

Background

Segmentation of gliomas in multi-parametric (MP-)MR images is challenging due to their heterogeneous nature in terms of size, appearance and location. Manual tumor segmentation is a time-consuming task and clinical practice would benefit from (semi-) automated segmentation of the different tumor compartments.

Methods

We present a semi-automated framework for brain tumor segmentation based on non-negative matrix factorization (NMF) that does not require prior training of the method. L1-regularization is incorporated into the NMF objective function to promote spatial consistency and sparseness of the tissue abundance maps. The pathological sources are initialized through user-defined voxel selection. Knowledge about the spatial location of the selected voxels is combined with tissue adjacency constraints in a post-processing step to enhance segmentation quality. The method is applied to an MP-MRI dataset of 21 high-grade glioma patients, including conventional, perfusion-weighted and diffusion-weighted MRI. To assess the effect of using MP-MRI data and the L1-regularization term, analyses are also run using only conventional MRI and without L1-regularization. Robustness against user input variability is verified by considering the statistical distribution of the segmentation results when repeatedly analyzing each patient’s dataset with a different set of random seeding points.

Results

Using L1-regularized semi-automated NMF segmentation, mean Dice-scores of 65%, 74 and 80% are found for active tumor, the tumor core and the whole tumor region. Mean Hausdorff distances of 6.1 mm, 7.4 mm and 8.2 mm are found for active tumor, the tumor core and the whole tumor region. Lower Dice-scores and higher Hausdorff distances are found without L1-regularization and when only considering conventional MRI data.

Conclusions

Based on the mean Dice-scores and Hausdorff distances, segmentation results are competitive with state-of-the-art in literature. Robust results were found for most patients, although careful voxel selection is mandatory to avoid sub-optimal segmentation.