Dice Similarity Index Research Articles

A simple rapid process for semi-automated brain extraction from magnetic resonance images of the whole mouse head

BackgroundMagnetic resonance imaging (MRI) is a well-developed technique in neuroscience. Limitations in applying MRI to rodent models of neuropsychiatric disorders include the large number of animals required to achieve statistical significance, and the paucity of automation tools for the critical early step in processing, brain extraction, which prepares brain images for alignment and voxel-wise statistics. New methodThis novel timesaving automation of template-based brain extraction (“skull-stripping”) is capable of quickly and reliably extracting the brain from large numbers of whole head images in a single step. The method is simple to install and requires minimal user interaction. ResultsThis method is equally applicable to different types of MR images. Results were evaluated with Dice and Jacquard similarity indices and compared in 3D surface projections with other stripping approaches. Statistical comparisons demonstrate that individual variation of brain volumes are preserved. Comparison with existing methodsA downloadable software package not otherwise available for extraction of brains from whole head images is included here. This software tool increases speed, can be used with an atlas or a template from within the dataset, and produces masks that need little further refinement. ConclusionsOur new automation can be applied to any MR dataset, since the starting point is a template mask generated specifically for that dataset. The method reliably and rapidly extracts brain images from whole head images, rendering them useable for subsequent analytical processing. This software tool will accelerate the exploitation of mouse models for the investigation of human brain disorders by MRI.

A novel approach for the automated segmentation and volume quantification of cardiac fats on computed tomography

The deposits of fat on the surroundings of the heart are correlated to several health risk factors such as atherosclerosis, carotid stiffness, coronary artery calcification, atrial fibrillation and many others. These deposits vary unrelated to obesity, which reinforces its direct segmentation for further quantification. However, manual segmentation of these fats has not been widely deployed in clinical practice due to the required human workload and consequential high cost of physicians and technicians. In this work, we propose a unified method for an autonomous segmentation and quantification of two types of cardiac fats. The segmented fats are termed epicardial and mediastinal, and stand apart from each other by the pericardium. Much effort was devoted to achieve minimal user intervention. The proposed methodology mainly comprises registration and classification algorithms to perform the desired segmentation. We compare the performance of several classification algorithms on this task, including neural networks, probabilistic models and decision tree algorithms. Experimental results of the proposed methodology have shown that the mean accuracy regarding both epicardial and mediastinal fats is 98.5% (99.5% if the features are normalized), with a mean true positive rate of 98.0%. In average, the Dice similarity index was equal to 97.6%.

Stratified mixture modeling for segmentation of white-matter lesions in brain MR images

Accurate characterization of white-matter lesions from magnetic resonance (MR) images has increasing importance for diagnosis and management of treatment of certain neurological diseases, and can be performed in an objective and effective way by automated lesion segmentation. This usually involves modeling the whole-brain MR intensity distribution, however, capturing various sources of MR intensity variability and lesion heterogeneity results in highly complex whole-brain MR intensity models, thus their robust estimation on a large set of MR images presents a huge challenge. We propose a novel approach employing stratified mixture modeling, where the main premise is that the otherwise complex whole-brain model can be reduced to a tractable parametric form in small brain subregions. We show on MR images of multiple sclerosis (MS) patients with different lesion loads that robust estimators enable accurate mixture modeling of MR intensity in small brain subregions even in the presence of lesions. Recombination of the mixture models across strata provided an accurate whole-brain MR intensity model. Increasing the number of subregions and, thereby, the model complexity, consistently improved the accuracy of whole-brain MR intensity modeling and segmentation of normal structures. The proposed approach was incorporated into three unsupervised lesion segmentation methods and, compared to original and three other state-of-the-art methods, the proposed modeling approach significantly improved lesion segmentation according to increased Dice similarity indices and lower number of false positives on real MR images of 30 patients with MS.

Automatic white matter lesion segmentation using contrast enhanced FLAIR intensity and Markov Random Field

White matter lesions (WMLs) are small groups of dead cells that clump together in the white matter of brain. In this paper, we propose a reliable method to automatically segment WMLs. Our method uses a novel filter to enhance the intensity of WMLs. Then a feature set containing enhanced intensity, anatomical and spatial information is used to train a random forest classifier for the initial segmentation of WMLs. Following that a reliable and robust edge potential function based Markov Random Field (MRF) is proposed to obtain the final segmentation by removing false positive WMLs. Quantitative evaluation of the proposed method is performed on 24 subjects of ENVISion study. The segmentation results are validated against the manual segmentation, performed under the supervision of an expert neuroradiologist. The results show a dice similarity index of 0.76 for severe lesion load, 0.73 for moderate lesion load and 0.61 for mild lesion load. In addition to that we have compared our method with three state of the art methods on 20 subjects of Medical Image Computing and Computer Aided Intervention Society's (MICCAI's) MS lesion challenge dataset, where our method shows better segmentation accuracy compare to the state of the art methods. These results indicate that the proposed method can assist the neuroradiologists in assessing the WMLs in clinical practice.

A method for the evaluation of thousands of automated 3D stem cell segmentations.

There is no segmentation method that performs perfectly with any dataset in comparison to human segmentation. Evaluation procedures for segmentation algorithms become critical for their selection. The problems associated with segmentation performance evaluations and visual verification of segmentation results are exaggerated when dealing with thousands of three-dimensional (3D) image volumes because of the amount of computation and manual inputs needed. We address the problem of evaluating 3D segmentation performance when segmentation is applied to thousands of confocal microscopy images (z-stacks). Our approach is to incorporate experimental imaging and geometrical criteria, and map them into computationally efficient segmentation algorithms that can be applied to a very large number of z-stacks. This is an alternative approach to considering existing segmentation methods and evaluating most state-of-the-art algorithms. We designed a methodology for 3D segmentation performance characterization that consists of design, evaluation and verification steps. The characterization integrates manual inputs from projected surrogate 'ground truth' of statistically representative samples and from visual inspection into the evaluation. The novelty of the methodology lies in (1) designing candidate segmentation algorithms by mapping imaging and geometrical criteria into algorithmic steps, and constructing plausible segmentation algorithms with respect to the order of algorithmic steps and their parameters, (2) evaluating segmentation accuracy using samples drawn from probability distribution estimates of candidate segmentations and (3) minimizing human labour needed to create surrogate 'truth' by approximating z-stack segmentations with 2D contours from three orthogonal z-stack projections and by developing visual verification tools. We demonstrate the methodology by applying it to a dataset of 1253 mesenchymal stem cells. The cells reside on 10 different types of biomaterial scaffolds, and are stained for actin and nucleus yielding 128 460 image frames (on average, 125 cells/scaffold × 10 scaffold types × 2 stains × 51 frames/cell). After constructing and evaluating six candidates of 3D segmentation algorithms, the most accurate 3D segmentation algorithm achieved an average precision of 0.82 and an accuracy of 0.84 as measured by the Dice similarity index where values greater than 0.7 indicate a good spatial overlap. A probability of segmentation success was 0.85 based on visual verification, and a computation time was 42.3 h to process all z-stacks. While the most accurate segmentation technique was 4.2 times slower than the second most accurate algorithm, it consumed on average 9.65 times less memory per z-stack segmentation.

Magnetic Resonance–Based Automatic Air Segmentation for Generation of Synthetic Computed Tomography Scans in the Head Region

To incorporate a novel imaging sequence for robust air and tissue segmentation using ultrashort echo time (UTE) phase images and to implement an innovative synthetic CT (synCT) solution as a first step toward MR-only radiation therapy treatment planning for brain cancer. Ten brain cancer patients were scanned with a UTE/Dixon sequence and other clinical sequences on a 1.0 T open magnet with simulation capabilities. Bone-enhanced images were generated from a weighted combination of water/fat maps derived from Dixon images and inverted UTE images. Automated air segmentation was performed using unwrapped UTE phase maps. Segmentation accuracy was assessed by calculating segmentation errors (true-positive rate, false-positive rate, and Dice similarity indices using CT simulation (CT-SIM) as ground truth. The synCTs were generated using a voxel-based, weighted summation method incorporating T2, fluid attenuated inversion recovery (FLAIR), UTE1, and bone-enhanced images. Mean absolute error (MAE) characterized Hounsfield unit (HU) differences between synCT and CT-SIM. A dosimetry study was conducted, and differences were quantified using γ-analysis and dose-volume histogram analysis. On average, true-positive rate and false-positive rate for the CT and MR-derived air masks were 80.8% ± 5.5% and 25.7% ± 6.9%, respectively. Dice similarity indices values were 0.78 ± 0.04 (range, 0.70-0.83). Full field of view MAE between synCT and CT-SIM was 147.5 ± 8.3 HU (range, 138.3-166.2 HU), with the largest errors occurring at bone-air interfaces (MAE 422.5 ± 33.4 HU for bone and 294.53 ± 90.56 HU for air). Gamma analysis revealed pass rates of 99.4% ± 0.04%, with acceptable treatment plan quality for the cohort. A hybrid MRI phase/magnitude UTE image processing technique was introduced that significantly improved bone and air contrast in MRI. Segmented air masks and bone-enhanced images were integrated into our synCT pipeline for brain, and results agreed well with clinical CTs, thereby supporting MR-only radiation therapy treatment planning in the brain.

Automatic delineation of tumor volumes by co-segmentation of combined PET/MR data

Combined PET/MRI may be highly beneficial for radiotherapy treatment planning in terms of tumor delineation and characterization. To standardize tumor volume delineation, an automatic algorithm for the co-segmentation of head and neck (HN) tumors based on PET/MR data was developed. Ten HN patient datasets acquired in a combined PET/MR system were available for this study. The proposed algorithm uses both the anatomical T2-weighted MR and FDG-PET data. For both imaging modalities tumor probability maps were derived, assigning each voxel a probability of being cancerous based on its signal intensity. A combination of these maps was subsequently segmented using a threshold level set algorithm. To validate the method, tumor delineations from three radiation oncologists were available. Inter-observer variabilities and variabilities between the algorithm and each observer were quantified by means of the Dice similarity index and a distance measure. Inter-observer variabilities and variabilities between observers and algorithm were found to be comparable, suggesting that the proposed algorithm is adequate for PET/MR co-segmentation. Moreover, taking into account combined PET/MR data resulted in more consistent tumor delineations compared to MR information only.

The Relative Importance of Domain Applicability Metrics for Estimating Prediction Errors in QSAR Varies with Training Set Diversity.

In QSAR, a statistical model is generated from a training set of molecules (represented by chemical descriptors) and their biological activities (an "activity model"). The aim of the field of domain applicability (DA) is to estimate the uncertainty of prediction of a specific molecule on a specific activity model. A number of DA metrics have been proposed in the literature for this purpose. A quantitative model of the prediction uncertainty (an "error model") can be built using one or more of these metrics. A previous publication from our laboratory ( Sheridan , R. P. J. Chem. Inf. 2013 , 53 , 2837 - 2850 ) suggested that QSAR methods such as random forest could be used to build error models by fitting unsigned prediction errors against DA metrics. The QSAR paradigm contains two useful techniques: descriptor importance can determine which DA metrics are most useful, and cross-validation can be used to tell which subset of DA metrics is sufficient to estimate the unsigned errors. Previously we studied 10 large, diverse data sets and seven DA metrics. For those data sets for which it is possible to build a significant error model from those seven metrics, only two metrics were sufficient to account for almost all of the information in the error model. These were TREE_SD (the variation of prediction among random forest trees) and PREDICTED (the predicted activity itself). In this paper we show that when data sets are less diverse, as for example in QSAR models of molecules in a single chemical series, these two DA metrics become less important in explaining prediction error, and the DA metric SIMILARITYNEAREST1 (the similarity of the molecule being predicted to the closest training set compound) becomes more important. Our recommendation is that when the mean pairwise similarity (measured with the Carhart AP descriptor and the Dice similarity index) within a QSAR training set is less than 0.5, one can use only TREE_SD, PREDICTED to form the error model, but otherwise one should use TREE_SD, PREDICTED, SIMILARITYNEAREST1.

SU-E-J-109: Evaluation of Deformable Accumulated Parotid Doses Using Different Registration Algorithms in Adaptive Head and Neck Radiotherapy

Purpose: Three deformable image registration (DIR) algorithms are utilized to perform deformable dose accumulation for head and neck tomotherapy treatment, and the differences of the accumulated doses are evaluated. Methods: Daily MVCT data for 10 patients with pathologically proven nasopharyngeal cancers were analyzed. The data were acquired using tomotherapy (TomoTherapy, Accuray) at the PLA General Hospital. The prescription dose to the primary target was 70Gy in 33 fractions.Three DIR methods (B-spline, Diffeomorphic Demons and MIMvista) were used to propagate parotid structures from planning CTs to the daily CTs and accumulate fractionated dose on the planning CTs. The mean accumulated doses of parotids were quantitatively compared and the uncertainties of the propagated parotid contours were evaluated using Dice similarity index (DSI). Results: The planned mean dose of the ipsilateral parotids (32.42±3.13Gy) was slightly higher than those of the contralateral parotids (31.38±3.19Gy)in 10 patients. The difference between the accumulated mean doses of the ipsilateral parotids in the B-spline, Demons and MIMvista deformation algorithms (36.40±5.78Gy, 34.08±6.72Gy and 33.72±2.63Gy ) were statistically significant (B-spline vs Demons, P<0.0001, B-spline vs MIMvista, p =0.002). And The difference between those of the contralateral parotids in the B-spline, Demons and MIMvista deformation algorithms (34.08±4.82Gy, 32.42±4.80Gy and 33.92±4.65Gy ) were also significant (B-spline vs Demons, p =0.009, B-spline vs MIMvista, p =0.074). For the DSI analysis, the scores of B-spline, Demons and MIMvista DIRs were 0.90, 0.89 and 0.76. Conclusion: Shrinkage of parotid volumes results in the dose increase to the parotid glands in adaptive head and neck radiotherapy. The accumulated doses of parotids show significant difference using the different DIR algorithms between kVCT and MVCT. Therefore, the volume-based criterion (i.e. DSI) as a quantitative evaluation of registration accuracy is essential besides the visual assessment by the treating physician. This work was supported in part by the grant from Chinese Natural Science Foundation (Grant No. 11105225).

TU‐CD‐BRA‐12: Coupling PET Image Restoration and Segmentation Using Variational Method with Multiple Regularizations

Purpose:To propose a new variational method which couples image restoration with tumor segmentation for PET images using multiple regularizations.Methods:Partial volume effect (PVE) is a major degrading factor impacting tumor segmentation accuracy in PET imaging. The existing segmentation methods usually need to take prior calibrations to compensate PVE and they are highly system‐dependent. Taking into account that image restoration and segmentation can promote each other and they are tightly coupled, we proposed a variational method to solve the two problems together. Our method integrated total variation (TV) semi‐blind deconvolution and Mumford‐Shah (MS) segmentation. The TV norm was used on edges to protect the edge information, and the L₂ norm was used to avoid staircase effect in the no‐edge area. The blur kernel was constrained to the Gaussian model parameterized by its variance and we assumed that the variances in the X‐Y and Z directions are different. The energy functional was iteratively optimized by an alternate minimization algorithm. Segmentation performance was tested on eleven patients with non‐Hodgkin's lymphoma, and evaluated by Dice similarity index (DSI) and classification error (CE). For comparison, seven other widely used methods were also tested and evaluated.Results:The combination of TV and L₂ regularizations effectively improved the segmentation accuracy. The average DSI increased by around 0.1 than using either the TV or the L₂ norm. The proposed method was obviously superior to other tested methods. It has an average DSI and CE of 0.80 and 0.41, while the FCM method — the second best one — has only an average DSI and CE of 0.66 and 0.64.Conclusion:Coupling image restoration and segmentation can handle PVE and thus improves tumor segmentation accuracy in PET. Alternate use of TV and L2 regularizations can further improve the performance of the algorithm.This work was supported in part by National Natural Science Foundation of China (NNSFC), under Grant No.61375018, and Fundamental Research Funds for the Central Universities, under Grant No. 2012QN086. Wei Lu was supported in part by the National Institutes of Health (NIH) Grant No. R01 CA172638.

Molecular characterization of an outbreak of carbapenem-resistant Klebsiella pneumoniae in a tertiary care hospital in Medellín, Colombia

Resistance to carbapenems is considered to represent a serious threat to public health at the global level, since these antibiotics are one of the last therapeutic options for the treatment of multidrug-resistant bacteria. Molecular characterization of outbreaks due to resistant bacteria provides information that can be used in the design of infection control strategies. To describe the molecular characteristics of an outbreak of carbapenem-resistant Klebsiella pneumoniae that occurred in a tertiary care hospital in Medellín in 2010-2011. Eighty-four isolates were obtained from a collection of strains associated with the hospital outbreak, of which 32 were from patients infected at that time and 52 were carriers. Identification and susceptibility of the isolates was performed using Vitek2®. Carbapenemases were detected using a modified Hodge test and polymerase chain reaction. Genetic relationships between the isolates were evaluated using pulsed field gel electrophoresis and multiple locus sequence typing. All the isolates analyzed were multidrug resistant; molecular analysis revealed that all harbored bla KPC-3 . The genetic analysis showed that 58/64 of the isolates from both infected and colonized patients were closely related (Dice similarity index >80%) and belonged to the ST258 lineage. Using molecular typing techniques it was possible to confirm the occurrence of an outbreak caused by K. pneumoniae ST258, a carrier of bla KPC-3 with a multidrug-resistant profile which had been associated with a previous outbreak in another hospital in the city of Medellín. ST258 is a high risk clone at the global level, demonstrating the potential for dissemination of resistance in this country. Implementation of molecular tools in support of epidemiological surveillance is useful for evaluating the spread of microorganisms of public health significance.

Neural-network based autocontouring algorithm for intrafractional lung-tumor tracking using Linac-MR.

To develop a neural-network based autocontouring algorithm for intrafractional lung-tumor tracking using Linac-MR and evaluate its performance with phantom and in-vivo MR images. An autocontouring algorithm was developed to determine both the shape and position of a lung tumor from each intrafractional MR image. A pulse-coupled neural network was implemented in the algorithm for contrast improvement of the tumor region. Prior to treatment, to initiate the algorithm, an expert user needs to contour the tumor and its maximum anticipated range of motion in pretreatment MR images. During treatment, however, the algorithm processes each intrafractional MR image and automatically generates a tumor contour without further user input. The algorithm is designed to produce a tumor contour that is the most similar to the expert's manual one. To evaluate the autocontouring algorithm in the author's Linac-MR environment which utilizes a 0.5 T MRI, a motion phantom and four lung cancer patients were imaged with 3 T MRI during normal breathing, and the image noise was degraded to reflect the image noise at 0.5 T. Each of the pseudo-0.5 T images was autocontoured using the author's algorithm. In each test image, the Dice similarity index (DSI) and Hausdorff distance (HD) between the expert's manual contour and the algorithm generated contour were calculated, and their centroid positions were compared (Δd centroid). The algorithm successfully contoured the shape of a moving tumor from dynamic MR images acquired every 275 ms. From the phantom study, mean DSI of 0.95-0.96, mean HD of 2.61-2.82 mm, and mean Δd centroid of 0.68-0.93 mm were achieved. From the in-vivo study, the author's algorithm achieved mean DSI of 0.87-0.92, mean HD of 3.12-4.35 mm, as well as Δd centroid of 1.03-1.35 mm. Autocontouring speed was less than 20 ms for each image. The authors have developed and evaluated a lung tumor autocontouring algorithm for intrafractional tumor tracking using Linac-MR. The autocontouring performance in the Linac-MR environment was evaluated using phantom and in-vivo MR images. From the in-vivo study, the author's algorithm achieved 87%-92% of contouring agreement and centroid tracking accuracy of 1.03-1.35 mm. These results demonstrate the feasibility of lung tumor autocontouring in the author's laboratory's Linac-MR environment.

Validated automatic brain extraction of head CT images

BackgroundX-ray computed tomography (CT) imaging of the brain is commonly used in diagnostic settings. Although CT scans are primarily used in clinical practice, they are increasingly used in research. A fundamental processing step in brain imaging research is brain extraction — the process of separating the brain tissue from all other tissues. Methods for brain extraction have either been 1) validated but not fully automated, or 2) fully automated and informally proposed, but never formally validated. AimTo systematically analyze and validate the performance of FSL's brain extraction tool (BET) on head CT images of patients with intracranial hemorrhage. This was done by comparing the manual gold standard with the results of several versions of automatic brain extraction and by estimating the reliability of automated segmentation of longitudinal scans. The effects of the choice of BET parameters and data smoothing is studied and reported. MethodsAll images were thresholded using a 0–100Hounsfield unit (HU) range. In one variant of the pipeline, data were smoothed using a 3-dimensional Gaussian kernel (σ=1mm3) and re-thresholded to 0–100HU; in the other, data were not smoothed. BET was applied using 1 of 3 fractional intensity (FI) thresholds: 0.01, 0.1, or 0.35 and any holes in the brain mask were filled.For validation against a manual segmentation, 36 images from patients with intracranial hemorrhage were selected from 19 different centers from the MISTIE (Minimally Invasive Surgery plus recombinant-tissue plasminogen activator for Intracerebral Evacuation) stroke trial. Intracranial masks of the brain were manually created by one expert CT reader. The resulting brain tissue masks were quantitatively compared to the manual segmentations using sensitivity, specificity, accuracy, and the Dice Similarity Index (DSI). Brain extraction performance across smoothing and FI thresholds was compared using the Wilcoxon signed-rank test. The intracranial volume (ICV) of each scan was estimated by multiplying the number of voxels in the brain mask by the dimensions of each voxel for that scan. From this, we calculated the ICV ratio comparing manual and automated segmentation: ICVautomatedICVmanual.To estimate the performance in a large number of scans, brain masks were generated from the 6 BET pipelines for 1095 longitudinal scans from 129 patients. Failure rates were estimated from visual inspection. ICV of each scan was estimated and an intraclass correlation (ICC) was estimated using a one-way ANOVA. ResultsSmoothing images improves brain extraction results using BET for all measures except specificity (all p<0.01, uncorrected), irrespective of the FI threshold. Using an FI of 0.01 or 0.1 performed better than 0.35. Thus, all reported results refer only to smoothed data using an FI of 0.01 or 0.1. Using an FI of 0.01 had a higher median sensitivity (0.9901) than an FI of 0.1 (0.9884, median difference: 0.0014, p<0.001), accuracy (0.9971 vs. 0.9971; median difference: 0.0001, p<0.001), and DSI (0.9895 vs. 0.9894; median difference: 0.0004, p<0.001) and lower specificity (0.9981 vs. 0.9982; median difference: −0.0001, p<0.001). These measures are all very high indicating that a range of FI values may produce visually indistinguishable brain extractions. Using smoothed data and an FI of 0.01, the mean (SD) ICV ratio was 1.002 (0.008); the mean being close to 1 indicates the ICV estimates are similar for automated and manual segmentation.In the 1095 longitudinal scans, this pipeline had a low failure rate (5.2%) and the ICC estimate was high (0.929, 95% CI: 0.91, 0.945) for successfully extracted brains. ConclusionBET performs well at brain extraction on thresholded, 1mm3 smoothed CT images with an FI of 0.01 or 0.1. Smoothing before applying BET is an important step not previously discussed in the literature. Analysis code is provided.

Disparate voxel based morphometry (VBM) results between SPM and FSL softwares in ALS patients with frontotemporal dementia: which VBM results to consider?

BackgroundBecause our previous study showed disparate voxel based morphometry (VBM) results between SPM and FSL softwares in the brain of amyotrophic lateral sclerosis patients with frontotemporal dementia (ALS-FTD), we investigated which VBM results may more represent atrophy by comparing with Freesurfer’s cortical volume and thickness measures.MethodsMRI at 1.5 T was obtained during routine clinical imaging of ALS-FTD patients (n = 18) and in unaffected neurologic controls (n = 15). Gray matter (GM) VBM analysis was carried out using FSL and SPM. Cortical thickness and volume analysis was performed using Freesurfer.ResultsGM volume was significantly (p < 0.05) reduced in both motor and extra motor regions in ALS- FTD when compared to unaffected neurologic controls in FSL and Freesurfer but not in SPM. Dice similarity index for cortical GM volume changes between FSL and Freesurfer was 0.30 for motor and 0.31 for non-motor regions as opposed to 0 (motor) and 0.02 (non-motor) between SPM and Freesurfer.ConclusionGM volume changes using FSL showed similar pattern with Freesurfer cortical volume and thickness changes in contrast to SPM results. Our results suggest that, at least for our dataset, VBM results obtained using FSL software should be considered as more representative of GM atrophy.

Abstract W P25: Anatomically-Weighted Predictive Algorithms Improve MRI-Based Tissue Outcome Predictions After Acute Ischemic Stroke

Background: In acute ischemic stroke (AIS) patients, multi-parametric MRI-based predictive algorithms have shown promise in identifying tissue at risk of infarction, but do not consider the intrinsic variations of normal or pathological tissue. We hypothesized that extending MRI-based algorithms to take into consideration tissue type will improve predictions of tissue outcome. Methods: We retrospectively analyzed AIS patients who received neither revascularization nor experimental interventional treatment, who underwent MRI within 12 h from the time since they were last known well and who had follow-up imaging &gt;4 days. Perfusion- and diffusion-parametric maps were combined to predict tissue outcome using 2 models: 1) a generalized linear model (GLM) trained with data from the whole ipsilateral hemisphere (sGLM), irrespective of tissue type, or 2) an anatomically-weighted GLM (aGLM) that was calculated using a weighted average to combine results from models generated using entire white or gray matter regions only. Both methods were evaluated using jack-knifing and predicted and follow-up regions were compared in terms of accuracy (measured as area under the receiver operator characteristic curve, AUC), Dice similarity index (DSI) and root mean square error (RMSE). Results: Results from 109 patients (65% male, median 68 y IQR [55-77], NIHSS 14 [9-25]) showed that, compared to sGLM, aGLM’s predictions had higher DSI (0.48 [0.19-0.59], P&lt;0.001), and AUC (0.89 [0.86-0.94], P=0.001) and lower RMSE (0.32 [0.29-0.35], P&lt;0.001), all demonstrating improved performance. Discussion: We showed that anatomically-weighted algorithms may better capture differences in tissue vulnerability in acute ischemic stroke, contributing to improved MRI-based tissue outcome predictions.

Automated Localization of Multiple Pelvic Bone Structures on MRI

In this paper, we present a fully automated localization method for multiple pelvic bone structures on magnetic resonance images (MRI). Pelvic bone structures are at present identified manually on MRI to locate reference points for measurement and evaluation of pelvic organ prolapse (POP). Given that this is a time-consuming and subjective procedure, there is a need to localize pelvic bone structures automatically. However, bone structures are not easily differentiable from soft tissue on MRI as their pixel intensities tend to be very similar. In this paper, we present a model that combines support vector machines and nonlinear regression capturing global and local information to automatically identify the bounding boxes of bone structures on MRI. The model identifies the location of the pelvic bone structures by establishing the association between their relative locations and using local information such as texture features. Results show that the proposed method is able to locate the bone structures of interest accurately (dice similarity index >0.75) in 87-91% of the images. This research aims to enable accurate, consistent, and fully automated localization of bone structures on MRI to facilitate and improve the diagnosis of health conditions such as female POP.

3D segmentation of annulus fibrosus and nucleus pulposus from T2-weighted magnetic resonance images

Computational medicine aims at employing personalised computational models in diagnosis and treatment planning. The use of such models to help physicians in finding the best treatment for low back pain (LBP) is becoming popular. One of the challenges of creating such models is to derive patient-specific anatomical and tissue models of the lumbar intervertebral discs (IVDs), as a prior step. This article presents a segmentation scheme that obtains accurate results irrespective of the degree of IVD degeneration, including pathological discs with protrusion or herniation. The segmentation algorithm, employing a novel feature selector, iteratively deforms an initial shape, which is projected into a statistical shape model space at first and then, into a B-Spline space to improve accuracy.The method was tested on a MR dataset of 59 patients suffering from LBP. The images follow a standard T2-weighted protocol in coronal and sagittal acquisitions. These two image volumes were fused in order to overcome large inter-slice spacing. The agreement between expert-delineated structures, used here as gold-standard, and our automatic segmentation was evaluated using Dice Similarity Index and surface-to-surface distances, obtaining a mean error of 0.68 mm in the annulus segmentation and 1.88 mm in the nucleus, which are the best results with respect to the image resolution in the current literature.

Sensitivity study of voxel-based PET image comparison to image registration algorithms.

Accurate deformable registration is essential for voxel-based comparison of sequential positron emission tomography (PET) images for proper adaptation of treatment plan and treatment response assessment. The comparison may be sensitive to the method of deformable registration as the optimal algorithm is unknown. This study investigated the impact of registration algorithm choice on therapy response evaluation. Sixteen patients with 20 lung tumors underwent a pre- and post-treatment computed tomography (CT) and 4D FDG-PET scans before and after chemoradiotherapy. All CT images were coregistered using a rigid and ten deformable registration algorithms. The resulting transformations were then applied to the respective PET images. Moreover, the tumor region defined by a physician on the registered PET images was classified into progressor, stable-disease, and responder subvolumes. Particularly, voxels with standardized uptake value (SUV) decreases >30% were classified as responder, while voxels with SUV increases >30% were progressor. All other voxels were considered stable-disease. The agreement of the subvolumes resulting from difference registration algorithms was assessed by Dice similarity index (DSI). Coefficient of variation (CV) was computed to assess variability of DSI between individual tumors. Root mean square difference (RMSrigid) of the rigidly registered CT images was used to measure the degree of tumor deformation. RMSrigid and DSI were correlated by Spearman correlation coefficient (R) to investigate the effect of tumor deformation on DSI. Median DSIrigid was found to be 72%, 66%, and 80%, for progressor, stable-disease, and responder, respectively. Median DSIdeformable was 63%-84%, 65%-81%, and 82%-89%. Variability of DSI was substantial and similar for both rigid and deformable algorithms with CV > 10% for all subvolumes. Tumor deformation had moderate to significant impact on DSI for progressor subvolume with Rrigid = - 0.60 (p = 0.01) and Rdeformable = - 0.46 (p = 0.01-0.20) averaging over all deformable algorithms. For stable-disease subvolumes, the correlations were significant (p < 0.001) for all registration algorithms with Rrigid = - 0.71 and Rdeformable = - 0.72. Progressor and stable-disease subvolumes resulting from rigid registration were in excellent agreement (DSI > 70%) for RMSrigid < 150 HU. However, tumor deformation was observed to have negligible effect on DSI for responder subvolumes with insignificant |R| < 0.26, p > 0.27. This study demonstrated that deformable algorithms cannot be arbitrarily chosen; different deformable algorithms can result in large differences of voxel-based PET image comparison. For low tumor deformation (RMSrigid < 150 HU), rigid and deformable algorithms yield similar results, suggesting deformable registration is not required for these cases.

Regression Segmentation for &lt;formula formulatype="inline"&gt;&lt;tex Notation="TeX"&gt;$M^{3}$&lt;/tex&gt;&lt;/formula&gt; Spinal Images

Clinical routine often requires to analyze spinal images of multiple anatomic structures in multiple anatomic planes from multiple imaging modalities (M(3)). Unfortunately, existing methods for segmenting spinal images are still limited to one specific structure, in one specific plane or from one specific modality (S(3)). In this paper, we propose a novel approach, Regression Segmentation, that is for the first time able to segment M(3) spinal images in one single unified framework. This approach formulates the segmentation task innovatively as a boundary regression problem: modeling a highly nonlinear mapping function from substantially diverse M(3) images directly to desired object boundaries. Leveraging the advancement of sparse kernel machines, regression segmentation is fulfilled by a multi-dimensional support vector regressor (MSVR) which operates in an implicit, high dimensional feature space where M(3) diversity and specificity can be systematically categorized, extracted, and handled. The proposed regression segmentation approach was thoroughly tested on images from 113 clinical subjects including both disc and vertebral structures, in both sagittal and axial planes, and from both MRI and CT modalities. The overall result reaches a high dice similarity index (DSI) 0.912 and a low boundary distance (BD) 0.928 mm. With our unified and expendable framework, an efficient clinical tool for M(3) spinal image segmentation can be easily achieved, and will substantially benefit the diagnosis and treatment of spinal diseases.

Segmentation of tumor and edema along with healthy tissues of brain using wavelets and neural networks.

Robust brain magnetic resonance (MR) segmentation algorithms are critical to analyze tissues and diagnose tumor and edema in a quantitative way. In this study, we present a new tissue segmentation algorithm that segments brain MR images into tumor, edema, white matter (WM), gray matter (GM), and cerebrospinal fluid (CSF). The detection of the healthy tissues is performed simultaneously with the diseased tissues because examining the change caused by the spread of tumor and edema on healthy tissues is very important for treatment planning. We used T1, T2, and FLAIR MR images of 20 subjects suffering from glial tumor. We developed an algorithm for stripping the skull before the segmentation process. The segmentation is performed using self-organizing map (SOM) that is trained with unsupervised learning algorithm and fine-tuned with learning vector quantization (LVQ). Unlike other studies, we developed an algorithm for clustering the SOM instead of using an additional network. Input feature vector is constructed with the features obtained from stationary wavelet transform (SWT) coefficients. The results showed that average dice similarity indexes are 91% for WM, 87% for GM, 96% for CSF, 61% for tumor, and 77% for edema.