Deep learning, data ramping, and uncertainty estimation for detecting artifacts in large, imbalanced databases of MRI images.

Magnetic resonance imaging (MRI) is increasingly being used to delineate morphological changes underlying neurological disorders. Successfully detecting these changes depends on the MRI data quality. Unfortunately, image artifacts frequently compromise the MRI utility, making it critical to screen the data. Currently, quality assessment requires visual inspection, a time-consuming process that suffers from inter-rater variability. Automated methods to detect MRI artifacts could improve the efficiency of the process. Such automated methods have achieved high accuracy using small datasets, with balanced proportions of MRI data with and without artifacts. With the current trend towards big data in neuroimaging, there is a need for automated methods that achieve accurate detection in large and imbalanced datasets. Deep learning (DL) is the ideal MRI artifact detection algorithm for large neuroimaging databases. However, the inference generated by DL does not commonly include a measure of uncertainty. Here, we present the first stochastic DL algorithm to generate automated, high-performing MRI artifact detection implemented on a large and imbalanced neuroimaging database. We implemented Monte Carlo dropout in a 3D AlexNet to generate probabilities and epistemic uncertainties. We then developed a method to handle class imbalance, namely data-ramping to transfer the learning by extending the dataset size and the proportion of the artifact-free data instances. We used a 34,800 scans (98% clean) dataset. At baseline, we obtained 89.3% testing accuracy (F1 = 0.230). Following the transfer learning (with data-ramping), we obtained 94.9% testing accuracy (F1 = 0.357) outperforming focal cross-entropy (92.9% testing accuracy, F1 = 0.304) incorporated for comparison at handling class imbalance. By implementing epistemic uncertainties, we improved the testing accuracy to 99.5% (F1 = 0.834), outperforming the results obtained in previous comparable studies. In addition, we estimated aleatoric uncertainties by incorporating random flips to the MRI volumes, and demonstrated that aleatoric uncertainty can be implemented as part of the pipeline. The methods we introduce enhance the efficiency of managing large databases and the exclusion of artifact images from big data analyses.

Cortical lesion hotspots and association of subpial lesions with disability in multiple sclerosis.

BACKGROUND: Dramatic improvements in visualization of cortical (especially subpial) multiple sclerosis (MS) lesions allow assessment of impact on clinical course. OBJECTIVE: Characterize cortical lesions by 7 tesla (T) T2*-/T1-weighted magnetic resonance imaging (MRI); determine relationship with other MS pathology and contribution to disability. METHODS: Sixty-four adults with MS (45 relapsing-remitting/19 progressive) underwent 3 T brain/spine MRI, 7 T brain MRI, and clinical testing. RESULTS: Cortical lesions were found in 94% (progressive: median 56/range 2-203; relapsing-remitting: 15/0-168; p = 0.004). Lesion distribution across 50 cortical regions was nonuniform (p = 0.006), with highest lesion burden in supplementary motor cortex and highest prevalence in superior frontal gyrus. Leukocortical and white matter lesion volumes were strongly correlated (r = 0.58, p < 0.0001), while subpial and white matter lesion volumes were moderately correlated (r = 0.30, p = 0.002). Leukocortical (p = 0.02) but not subpial lesions (p = 0.40) were correlated with paramagnetic rim lesions; both were correlated with spinal cord lesions (p = 0.01). Cortical lesion volumes (total and subtypes) were correlated with expanded disability status scale, 25-foot timed walk, nine-hole peg test, and symbol digit modality test scores. CONCLUSION: Cortical lesions are highly prevalent and are associated with disability and progressive disease. Subpial lesion burden is not strongly correlated with white matter lesions, suggesting differences in inflammation and repair mechanisms.

Using Deep Learning Algorithms to Automatically Identify the Brain MRI Contrast: Implications for Managing Large Databases.

Neuroimaging science has seen a recent explosion in dataset size driving the need to develop database management with efficient processing pipelines. Multi-center neuroimaging databases consistently receive magnetic resonance imaging (MRI) data with unlabeled or incorrectly labeled contrast. There is a need to automatically identify the contrast of MRI scans to save database-managing facilities valuable resources spent by trained technicians required for visual inspection. We developed a deep learning (DL) algorithm with convolution neural network architecture to automatically infer the contrast of MRI scans based on the image intensity of multiple slices. For comparison, we developed a random forest (RF) algorithm to automatically infer the contrast of MRI scans based on acquisition parameters. The DL algorithm was able to automatically identify the MRI contrast of an unseen dataset with <0.2% error rate. The RF algorithm was able to identify the MRI contrast of the same dataset with 1.74% error rate. Our analysis showed that reduced dataset sizes caused the DL algorithm to lose generalizability. Finally, we developed a confidence measure, which made it possible to detect, with 100% specificity, all MRI volumes that were misclassified by the DL algorithm. This confidence measure can be used to alert the user on the need to inspect the small fraction of MRI volumes that are prone to misclassification. Our study introduces a practical solution for automatically identifying the MRI contrast. Furthermore, it demonstrates the powerful combination of convolution neural networks and DL for analyzing large MRI datasets.

Neuroprotection with glatiramer acetate: evidence from the PreCISe trial.

The phase III, multicenter, randomized, placebo-controlled PreCISe trial assessed glatiramer acetate (GA) effects in patients with clinically isolated syndromes (CIS) suggestive of multiple sclerosis (MS). To assess the neuroprotective effect of GA in a subset of patients in the PreCISe trial, we used proton magnetic resonance spectroscopy (MRS) to measure N-acetylaspartate (NAA), a marker of neuronal integrity, in a large central volume of brain. Thirty-four CIS patients randomized to GA 20 mg/day (n = 19) SC or placebo (n = 15) were included. Patients who relapsed (developed clinically definite MS [CDMS]) were removed from the substudy. NAA/creatine (NAA/Cr) ratios were compared between GA-treated and placebo-treated patients. Twenty patients with CIS had not converted to CDMS and were still in the double-blind phase of the trial at 12 months of follow-up. Paired changes in NAA/Cr differed significantly in patients treated with GA (+0.14, n = 11) compared with patients receiving placebo (-0.33, n = 9, p = 0.03) at 12 months, consistent with a neuroprotective effect of GA in vivo. Patients with CIS who received GA showed improvement in brain neuroaxonal integrity, as indicated by increased NAA/Cr, relative to comparable patients treated with placebo, who showed a decline in NAA/Cr consistent with findings from natural history studies.

Primary progressive multiple sclerosis: part of the MS disease spectrum or separate disease entity?

Multiple sclerosis (MS), the most frequent demyelinating disease, is characterized by a variable disease course. The majority of patients starts with relapsing remitting (RR) disease; approximately 50-60% of these patients progress to secondary progressive (SP) disease. Only about 15% of the patients develop a progressive disease course from onset, termed primary progressive multiple sclerosis (PPMS); the underlying pathogenic mechanisms responsible for onset of the disease with either PPMS or relapsing remitting multiple sclerosis (RRMS) are unknown. Patients with PPMS do not show a female predominance and usually have a later onset of disease compared to patients with RRMS. Monozygous twins can be concordant or discordant for disease courses indicating that the disease course is not only genetically determined. Primary progressive multiple sclerosis and secondary progressive multiple sclerosis (SPMS) share many similarities in imaging and pathological findings. Differences observed among the different disease courses are more of a quantitative than qualitative nature suggesting that the different phenotypes are part of a disease spectrum modulated by individual genetic predisposition and environmental influences. In this review, we summarize the knowledge regarding the clinical, epidemiological, imaging, and pathological characteristics of PPMS and compare those characteristics with RRMS and SPMS.

The metabolic epicenter of supratentorial gliomas: a 1H-MRSI study.

BACKGROUND: Assessing the impact of glioma location on prognosis remains elusive. We approached the problem using multivoxel proton magnetic resonance spectroscopic imaging (1H-MRSI) to define a tumor "metabolic epicenter", and examined the relationship of metabolic epicenter location to survival and histopathological grade. METHODS: We studied 54 consecutive patients with a supratentorial glioma (astrocytoma or oligodendroglioma, WHO grades II-IV). The metabolic epicenter in each tumor was defined as the 1H-MRSI voxel containing maximum intra-tumoral choline on preoperative imaging. Tumor location was considered the X-Y-Z coordinate position, in a standardized stereotactic space, of the metabolic epicenter. Correlation between epicenter location and survival or grade was assessed. RESULTS: Metabolic epicenter location correlated significantly with patient survival for all tumors (r2 = 0.30, p = 0.0002) and astrocytomas alone (r2 = 0.32, p = 0.005). A predictive model based on both metabolic epicenter location and histopathological grade accounted for 70% of the variability in survival, substantially improving on histology alone to predict survival. Location also correlated significantly with grade (r2 = 0.25, p = 0.001): higher grade tumors had a metabolic epicenter closer to the midpoint of the brain. CONCLUSIONS: The concept of the metabolic epicenter eliminates several problems related to existing methods of classifying glioma location. The location of the metabolic epicenter is strongly correlated with overall survival and histopathological grade, suggesting that it reflects biological factors underlying glioma growth and malignant dedifferentiation. These findings may be clinically relevant to predicting patterns of local glioma recurrence, and in planning resective surgery or radiotherapy.

Lesion distribution in children with clinically isolated syndromes.

We evaluated the incidence, volume, and spatial distribution of T2-weighted magnetic resonance imaging lesions in 58 children with clinically isolated syndromes at risk for multiple sclerosis compared with 58 adults with relapsing-remitting multiple sclerosis. Pediatric patients with clinically isolated syndromes who had brain lesions had supratentorial lesion volumes similar to adult multiple sclerosis patients, but greater infratentorial lesion volumes (p < 0.009), particularly in the pons of male patients. The predilection for infratentorial lesions the pediatric patients with clinically isolated syndromes may reflect immunological differences or differences in myelin, possibly related to the caudorostral temporal gradient in myelin maturation.

In vivo profiling of focal cortical dysplasia on high-resolution MRI with computational models.

PURPOSE: On MRI, focal cortical dysplasia (FCD) is characterized by a combination of increased cortical thickness, hyperintense signal within the dysplastic lesion, and blurred transition between gray and white matter (GM-WM). The visual identification of these abnormal characteristics may be difficult, and it is unclear to what degree these features occur among different FCD lesions. Our purpose was to investigate the pattern of occurrence of abnormal MRI characteristics in FCD by using a set of computational models and to generate quantitative lesion profiling. METHODS: A set of voxel-wise operators was applied to high-resolution 3D T1-weighted MRI in 23 patients with histologically proven FCD and 39 healthy controls, creating maps of GM thickness, maps of relative intensity highlighting areas with hyperintense signal, and maps of gradient magnitude modeling the GM-WM transition. All FCD lesions were segmented manually on the T1-weighted MRI. RESULTS: FCD volumes ranged from 734 mm3 to 80,726 mm3 (mean, 8,629 mm3 +/- 16,238). The manually segmented FCD lesions were used to estimate features in the lesional area and to determine possible local variations of each feature by means of a histogram. In 78% of the patients, FCD lesions were characterized by simultaneous GM thickening, hyperintense signal, and blurring of the GM-WM transition. Moreover, in all patients, the FCD lesion had at least two of these three characteristics. CONCLUSIONS: The three features occurred regardless of the lesion volume, and they characterized not only large FCD lesions, but also subtle ones that had been overlooked by conventional radiologic inspection before surgery.

Axonal injury in the cerebral normal-appearing white matter of patients with multiple sclerosis is related to concurrent demyelination in lesions but not to concurrent demyelination in normal-appearing white matter.

We assessed axonal injury and demyelination in the cerebral normal-appearing white matter (NAWM) of MS patients in a pilot study using proton magnetic resonance spectroscopic imaging and quantitative magnetization transfer (MT) imaging. Resonance intensities of N-acetylaspartate (NAA) relative to creatine (Cr) were measured in a large central brain volume. NAA/Cr in NAWM was estimated by regression of the NAA/Cr in each voxel against white matter fraction and extrapolation to a white matter fraction of 1. The fractional size of the semi-solid pool (F) was obtained from the binary spin bath model of MT by computing the model parameters from multiple MT-weighted and relaxometry acquisitions. F in NAWM was significantly smaller in the patients [0.109 (0.009)] relative to controls [0.123 (0.007), P = 0.011], but did not differ between RR [0.1085] and SP [0.1087] patients [P > 0.99]. NAA/Cr and F in the NAWM were not correlated (r = 0.16, P > 0.7), mainly due to a lack of variation in F among patients. This may indicate a floor to the extent of myelin pathology that can occur in NAWM before a lesion appears, or that axonal damage is not strictly related to demyelination. The correlation between NAWM NAA/Cr and T2w lesion volume was not significant (P > 0.1). However, dividing the lesion volumes by the mean F in T2w lesions resulted in a quantity that correlated well with NAWM NAA/Cr (r = -0.78, P = 0.038), possibly reflecting the association of Wallerian degeneration in the NAWM with axonal transection associated with demyelination within lesions.

Automated detection of focal cortical dysplasia lesions using computational models of their MRI characteristics and texture analysis.

Focal cortical dysplasia (FCD), a malformation of cortical development, is a frequent cause of pharmacologically intractable epilepsy. FCD is characterized on Tl-weighted MRI by cortical thickening, blurring of the gray-matter/white-matter interface, and gray-level hyperintensity. We have previously used computational models of these characteristics to enhance visual lesion detection. In the present study we seek to improve our methods by combining these models with features derived from texture analysis of MRI, which allows measurement of image properties not readily accessible by visual analysis. These computational models and texture features were used to develop a two-stage Bayesian classifier to perform automated FCD lesion detection. Eighteen patients with histologically confirmed FCD and 14 normal controls were studied. On the MRI volumes of the 18 patients, 20 FCD lesions were manually labeled by an expert observer. Three-dimensional maps of the computational models and texture features were constructed for all subjects. A Bayesian classifier was trained on the computational models to classify voxels as cerebrospinal fluid, gray-matter, white-matter, transitional, or lesional. Voxels classified as lesional were subsequently reclassified based on the texture features. This process produced a 3D lesion map, which was compared to the manual lesion labels. The automated classifier identified 17/20 manually labeled lesions. No lesions were identified in controls. Thus, combining models of the T1-weighted MRI characteristics of FCD with texture analysis enabled successful construction of a classifier. This computer-based, automated method may be useful in the presurgical evaluation of patients with severe epilepsy related to FCD.

Proton magnetic resonance spectroscopic imaging can predict length of survival in patients with supratentorial gliomas.

OBJECTIVE: We compared the ability of proton magnetic resonance spectroscopic imaging ((1)H-MRSI) measures with that of standard clinicopathological measures to predict length of survival in patients with supratentorial gliomas. METHODS: We developed two sets of leave-one-out logistic regression models based on either 1) intratumoral (1)H-MRSI features, including maximum values of a) choline and b) lactate-lipid, c) number of (1)H-MRSI voxels with low N-acetyl group values, and d) number of (1)H-MRSI voxels with high lactate-lipid values, all (a-d) of which were normalized to creatine in normal-appearing brain, or 2) standard clinicopathological features, including a) tumor histopathological grade, b) patient age, c) performance of surgical debulking, and d) tumor diagnosis (i.e., oligodendroglioma, astrocytoma). We assessed the accuracy of these two models in predicting patient survival for 6, 12, 24, and 48 months by performing receiver operating characteristic curve analysis. Cox proportional hazards analysis was performed to assess the extent to which patient survival could be explained by the above predictors. We then performed a series of leave-one-out linear multiple regression analyses to determine how well patient survival could be predicted in a continuous fashion. RESULTS: The results of using the models based on (1)H-MRSI and clinicopathological features were equally good, accounting for 81 and 64% of the variability (r(2)) in patients' actual survival durations. All features except number of (1)H-MRSI voxels with lactate-lipid/creatine values of at least 1 were significant predictors of survival in the (1)H-MRSI model. Two features (tumor grade and debulking) were found to be significant predictors in the clinicopathological model. Survival as a continuous variable was predicted accurately on the basis of the (1)H-MRSI data (r = 0.77, P < 0.001; median prediction error, 1.7 mo). CONCLUSION: Our results suggest that appropriate analysis of (1)H-MRSI data can predict survival in patients with supratentorial gliomas at least as accurately as data derived from more invasive clinicopathological features.

Computational models of MRI characteristics of focal cortical dysplasia improve lesion detection.

In many patients, focal cortical dysplasia (FCD) is characterized by minor structural changes that may go unrecognized by standard radiological analysis. We previously demonstrated that visual analysis of a composite map based on three simple models of MRI features of FCD increased the sensitivity of FCD lesion detection, compared to visual analysis of conventional MRI. Here we report on the use of improved methods for characterizing FCD which improve contrast in the composite maps: a Laplacian-based metric for measuring cortical thickness, a convolutional kernel to model blurring of the GM-WM interface, and an operator to measure hyperintense T1 signal. To validate these methods, we processed the MRIs of 14 FCD patients with our original set of image processing operators and an improved set of image processing operators. Comparison of the composite maps associated with the two sets of operators revealed that contrast between lesional tissue and nonlesional cortex was significantly increased in the composite maps associated with the set of improved operators. Increasing this contrast is an important step toward the goal of automated FCD lesion detection.