The data representativeness criterion: Predicting the performance of supervised classification based on data set similarity.

In a broad range of fields it may be desirable to reuse a supervised classification algorithm and apply it to a new data set. However, generalization of such an algorithm and thus achieving a similar classification performance is only possible when the training data used to build the algorithm is similar to new unseen data one wishes to apply it to. It is often unknown in advance how an algorithm will perform on new unseen data, being a crucial reason for not deploying an algorithm at all. Therefore, tools are needed to measure the similarity of data sets. In this paper, we propose the Data Representativeness Criterion (DRC) to determine how representative a training data set is of a new unseen data set. We present a proof of principle, to see whether the DRC can quantify the similarity of data sets and whether the DRC relates to the performance of a supervised classification algorithm. We compared a number of magnetic resonance imaging (MRI) data sets, ranging from subtle to severe difference is acquisition parameters. Results indicate that, based on the similarity of data sets, the DRC is able to give an indication as to when the performance of a supervised classifier decreases. The strictness of the DRC can be set by the user, depending on what one considers to be an acceptable underperformance.

Evaluation of a deep learning approach for the segmentation of brain tissues and white matter hyperintensities of presumed vascular origin in MRI.

Automatic segmentation of brain tissues and white matter hyperintensities of presumed vascular origin (WMH) in MRI of older patients is widely described in the literature. Although brain abnormalities and motion artefacts are common in this age group, most segmentation methods are not evaluated in a setting that includes these items. In the present study, our tissue segmentation method for brain MRI was extended and evaluated for additional WMH segmentation. Furthermore, our method was evaluated in two large cohorts with a realistic variation in brain abnormalities and motion artefacts. The method uses a multi-scale convolutional neural network with a T1-weighted image, a T2-weighted fluid attenuated inversion recovery (FLAIR) image and a T1-weighted inversion recovery (IR) image as input. The method automatically segments white matter (WM), cortical grey matter (cGM), basal ganglia and thalami (BGT), cerebellum (CB), brain stem (BS), lateral ventricular cerebrospinal fluid (lvCSF), peripheral cerebrospinal fluid (pCSF), and WMH. Our method was evaluated quantitatively with images publicly available from the MRBrainS13 challenge (n = 20), quantitatively and qualitatively in relatively healthy older subjects (n = 96), and qualitatively in patients from a memory clinic (n = 110). The method can accurately segment WMH (Overall Dice coefficient in the MRBrainS13 data of 0.67) without compromising performance for tissue segmentations (Overall Dice coefficients in the MRBrainS13 data of 0.87 for WM, 0.85 for cGM, 0.82 for BGT, 0.93 for CB, 0.92 for BS, 0.93 for lvCSF, 0.76 for pCSF). Furthermore, the automatic WMH volumes showed a high correlation with manual WMH volumes (Spearman's ρ = 0.83 for relatively healthy older subjects). In both cohorts, our method produced reliable segmentations (as determined by a human observer) in most images (relatively healthy/memory clinic: tissues 88%/77% reliable, WMH 85%/84% reliable) despite various degrees of brain abnormalities and motion artefacts. In conclusion, this study shows that a convolutional neural network-based segmentation method can accurately segment brain tissues and WMH in MR images of older patients with varying degrees of brain abnormalities and motion artefacts.

Robustness of Automated Methods for Brain Volume Measurements across Different MRI Field Strengths.

INTRODUCTION: Pooling of multicenter brain imaging data is a trend in studies on ageing related brain diseases. This poses challenges to MR-based brain segmentation. The performance across different field strengths of three widely used automated methods for brain volume measurements was assessed in the present study. METHODS: Ten subjects (mean age: 64 years) were scanned on 1.5T and 3T MRI on the same day. We determined robustness across field strength (i.e., whether measured volumes between 3T and 1.5T scans in the same subjects were similar) for SPM12, Freesurfer 5.3.0 and FSL 5.0.7. As a frame of reference, 3T MRI scans from 20 additional subjects (mean age: 71 years) were segmented manually to determine accuracy of the methods (i.e., whether measured volumes corresponded with expert-defined volumes). RESULTS: Total brain volume (TBV) measurements were robust across field strength for Freesurfer and FSL (mean absolute difference as % of mean volume ≤ 1%), but less so for SPM (4%). Gray matter (GM) and white matter (WM) volume measurements were robust for Freesurfer (1%; 2%) and FSL (2%; 3%) but less so for SPM (5%; 4%). For intracranial volume (ICV), SPM was more robust (2%) than FSL (3%) and Freesurfer (9%). TBV measurements were accurate for SPM and FSL, but less so for Freesurfer. For GM volume, SPM was accurate, but accuracy was lower for Freesurfer and FSL. For WM volume, Freesurfer was accurate, but SPM and FSL were less accurate. For ICV, FSL was accurate, while SPM and Freesurfer were less accurate. CONCLUSION: Brain volumes and ICV could be measured quite robustly in scans acquired at different field strengths, but performance of the methods varied depending on the assessed compartment (e.g., TBV or ICV). Selection of an appropriate method in multicenter brain imaging studies therefore depends on the compartment of interest.

Automatic brain tissue segmentation in MR images using Random Forests and Conditional Random Fields.

BACKGROUND: The segmentation of brain tissue into cerebrospinal fluid, gray matter, and white matter in magnetic resonance imaging scans is an important procedure to extract regions of interest for quantitative analysis and disease assessment. Manual segmentation requires skilled experts, being a laborious and time-consuming task; therefore, reliable and robust automatic segmentation methods are necessary. NEW METHOD: We propose a segmentation framework based on a Conditional Random Field for brain tissue segmentation, with a Random Forest encoding the likelihood function. The features include intensities, gradients, probability maps, and locations. Additionally, skull stripping is critical for achieving an accurate segmentation; thus, after extracting the brain we propose to refine its boundary during segmentation. RESULTS: The proposed framework was evaluated on the MR Brain Image Segmentation Challenge and the Internet Brain Segmentation Repository databases. The segmentations of brain tissues obtained with the proposed algorithm were competitive both in normal and diseased subjects. The skull stripping refinement significantly improved the results, when comparing against no refinement. COMPARISON WITH EXISTING METHODS: In the MR Brain Image Segmentation Challenge database, the results were competitive when comparing with top methods. In the Internet Brain Segmentation Repository database, the proposed approach outperformed other well-established algorithms. CONCLUSIONS: The combination of a Random Forest and Conditional Random Field for brain tissue segmentation performed well for normal and diseased subjects. Additionally, refinement of the skull stripping at segmentation time is feasible in learning-based methods and significantly improves the segmentation of cerebrospinal fluid and intracranial volume.

Automatic Segmentation of MR Brain Images With a Convolutional Neural Network.

Automatic segmentation in MR brain images is important for quantitative analysis in large-scale studies with images acquired at all ages. This paper presents a method for the automatic segmentation of MR brain images into a number of tissue classes using a convolutional neural network. To ensure that the method obtains accurate segmentation details as well as spatial consistency, the network uses multiple patch sizes and multiple convolution kernel sizes to acquire multi-scale information about each voxel. The method is not dependent on explicit features, but learns to recognise the information that is important for the classification based on training data. The method requires a single anatomical MR image only. The segmentation method is applied to five different data sets: coronal T2-weighted images of preterm infants acquired at 30 weeks postmenstrual age (PMA) and 40 weeks PMA, axial T2-weighted images of preterm infants acquired at 40 weeks PMA, axial T1-weighted images of ageing adults acquired at an average age of 70 years, and T1-weighted images of young adults acquired at an average age of 23 years. The method obtained the following average Dice coefficients over all segmented tissue classes for each data set, respectively: 0.87, 0.82, 0.84, 0.86, and 0.91. The results demonstrate that the method obtains accurate segmentations in all five sets, and hence demonstrates its robustness to differences in age and acquisition protocol.

MRBrainS Challenge: Online Evaluation Framework for Brain Image Segmentation in 3T MRI Scans.

Many methods have been proposed for tissue segmentation in brain MRI scans. The multitude of methods proposed complicates the choice of one method above others. We have therefore established the MRBrainS online evaluation framework for evaluating (semi)automatic algorithms that segment gray matter (GM), white matter (WM), and cerebrospinal fluid (CSF) on 3T brain MRI scans of elderly subjects (65-80 y). Participants apply their algorithms to the provided data, after which their results are evaluated and ranked. Full manual segmentations of GM, WM, and CSF are available for all scans and used as the reference standard. Five datasets are provided for training and fifteen for testing. The evaluated methods are ranked based on their overall performance to segment GM, WM, and CSF and evaluated using three evaluation metrics (Dice, H95, and AVD) and the results are published on the MRBrainS13 website. We present the results of eleven segmentation algorithms that participated in the MRBrainS13 challenge workshop at MICCAI, where the framework was launched, and three commonly used freeware packages: FreeSurfer, FSL, and SPM. The MRBrainS evaluation framework provides an objective and direct comparison of all evaluated algorithms and can aid in selecting the best performing method for the segmentation goal at hand.

Total bolus extraction method improves arterial image quality in dynamic CTAs derived from whole-brain CTP data.

Background and Purposes. The 320-detector row CT scanner enables visualization of whole-brain hemodynamic information (dynamic CT angiography (CTA) derived from CT perfusion scans). However, arterial image quality in dynamic CTA (dCTA) is inferior to arterial image quality in standard CTA. This study evaluates whether the arterial image quality can be improved by using a total bolus extraction (ToBE) method. Materials and Methods. DCTAs of 15 patients, who presented with signs of acute cerebral ischemia, were derived from 320-slice CT perfusion scans using both the standard subtraction method and the proposed ToBE method. Two neurointerventionalists blinded to the scan type scored the arterial image quality on a 5-point scale in the 4D dCTAs in consensus. Arteries were divided into four categories: (I) large extradural, (II) intradural (large, medium, and small), (III) communicating arteries, and (IV) cerebellar and ophthalmic arteries. Results. Quality of extradural and intradural arteries was significantly higher in the ToBE dCTAs than in the standard dCTAs (extradural P = 0.001, large intradural P < 0.001, medium intradural P < 0.001, and small intradural P < 0.001). Conclusion. The 4D dCTAs derived with the total bolus extraction (ToBE) method provide hemodynamic information combined with improved arterial image quality as compared to standard 4D dCTAs.

Timing-invariant reconstruction for deriving high-quality CT angiographic data from cerebral CT perfusion data.

PURPOSE: To suggest a simple and robust technique used to reconstruct high-quality computed tomographic (CT) angiographic images from CT perfusion data and to compare it with currently used CT angiography techniques. MATERIALS AND METHODS: Institutional review board approval was waived for this retrospective study, which included 25 consecutive patients who had had a stroke. Temporal maximum intensity projection (tMIP) CT angiographic images were created by using prior temporal filtering as a timing-insensitive technique to produce CT angiographic images from CT perfusion data. The temporal filter strength was optimized to gain maximal contrast-to-noise ratios (CNRs) in the circle of Willis. The resulting timing-invariant (TI) CT angiography was compared with standard helical CT angiography, the arterial phase of dynamic CT angiography, and nonfiltered tMIP CT angiography. Vascular contrast, image noise, and CNR were measured. Four experienced observers scored all images for vascular noise, vascular contour, detail of small and medium arteries, venous superimposition, and overall image quality in a blinded side-by-side comparison. Measurements were compared with a paired t test; P ≤ .05 indicated a significant difference. RESULTS: On average, optimized temporal filtering in TI CT angiography increased CNR by 18% and decreased image noise by 18% at the expense of a decrease in vascular contrast of 3% when compared with nonfiltered tMIP CT angiography. CNR, image noise, vascular noise, vascular contour, detail visibility of small and medium arteries, and overall image quality of TI CT angiograms were superior to those of standard CT angiography, tMIP CT angiography, and the arterial phase of dynamic CT angiography at a vascular contrast that was similar to that of standard CT angiography. Venous superimposition was similar for all techniques. Image quality of the arterial phase of dynamic CT angiography was rated inferior to that of standard CT angiography. CONCLUSION: TI CT angiographic images constructed by using temporally filtered tMIP CT angiographic data have excellent image quality that is superior to that achieved with currently used techniques, but they suffer from modest venous superimposition.

TIPS bilateral noise reduction in 4D CT perfusion scans produces high-quality cerebral blood flow maps.

Cerebral computed tomography perfusion (CTP) scans are acquired to detect areas of abnormal perfusion in patients with cerebrovascular diseases. These 4D CTP scans consist of multiple sequential 3D CT scans over time. Therefore, to reduce radiation exposure to the patient, the amount of x-ray radiation that can be used per sequential scan is limited, which results in a high level of noise. To detect areas of abnormal perfusion, perfusion parameters are derived from the CTP data, such as the cerebral blood flow (CBF). Algorithms to determine perfusion parameters, especially singular value decomposition, are very sensitive to noise. Therefore, noise reduction is an important preprocessing step for CTP analysis. In this paper, we propose a time-intensity profile similarity (TIPS) bilateral filter to reduce noise in 4D CTP scans, while preserving the time-intensity profiles (fourth dimension) that are essential for determining the perfusion parameters. The proposed TIPS bilateral filter is compared to standard Gaussian filtering, and 4D and 3D (applied separately to each sequential scan) bilateral filtering on both phantom and patient data. Results on the phantom data show that the TIPS bilateral filter is best able to approach the ground truth (noise-free phantom), compared to the other filtering methods (lowest root mean square error). An observer study is performed using CBF maps derived from fifteen CTP scans of acute stroke patients filtered with standard Gaussian, 3D, 4D and TIPS bilateral filtering. These CBF maps were blindly presented to two observers that indicated which map they preferred for (1) gray/white matter differentiation, (2) detectability of infarcted area and (3) overall image quality. Based on these results, the TIPS bilateral filter ranked best and its CBF maps were scored to have the best overall image quality in 100% of the cases by both observers. Furthermore, quantitative CBF and cerebral blood volume values in both the phantom and the patient data showed that the TIPS bilateral filter resulted in realistic mean values with a smaller standard deviation than the other evaluated filters and higher contrast-to-noise ratios. Therefore, applying the proposed TIPS bilateral filtering method to 4D CTP data produces higher quality CBF maps than applying the standard Gaussian, 3D bilateral or 4D bilateral filter. Furthermore, the TIPS bilateral filter is computationally faster than both the 3D and 4D bilateral filters.

Noise reduction in computed tomography scans using 3-d anisotropic hybrid diffusion with continuous switch.

Noise filtering techniques that maintain image contrast while decreasing image noise have the potential to optimize the quality of computed tomography (CT) images acquired at reduced radiation dose. In this paper, a hybrid diffusion filter with continuous switch (HDCS) is introduced, which exploits the benefits of three-dimensional edge-enhancing diffusion (EED) and coherence-enhancing diffusion (CED). Noise is filtered, while edges, tubular structures, and small spherical structures are preserved. From ten high dose thorax CT scans, acquired at clinical doses, ultra low dose ( 15 mAs ) scans were simulated and used to evaluate and compare HDCS to other diffusion filters, such as regularized Perona-Malik diffusion and EED. Quantitative results show that the HDCS filter outperforms the other filters in restoring the high dose CT scan from the corresponding simulated low dose scan. A qualitative evaluation was performed on filtered real low dose CT thorax scans. An expert observer scored artifacts as well as fine structures and was asked to choose one of three scans (two filtered (blinded), one unfiltered) for three different settings (trachea, lung, and mediastinal). Overall, the HDCS filtered scan was chosen most often.