Voxel Classification Research Articles

Automatic coronary artery calcium scoring in cardiac CT angiography using paired convolutional neural networks.

The amount of coronary artery calcification (CAC) is a strong and independent predictor of cardiovascular events. CAC is clinically quantified in cardiac calcium scoring CT (CSCT), but it has been shown that cardiac CT angiography (CCTA) may also be used for this purpose. We present a method for automatic CAC quantification in CCTA. This method uses supervised learning to directly identify and quantify CAC without a need for coronary artery extraction commonly used in existing methods. The study included cardiac CT exams of 250 patients for whom both a CCTA and a CSCT scan were available. To restrict the volume-of-interest for analysis, a bounding box around the heart is automatically determined. The bounding box detection algorithm employs a combination of three ConvNets, where each detects the heart in a different orthogonal plane (axial, sagittal, coronal). These ConvNets were trained using 50 cardiac CT exams. In the remaining 200 exams, a reference standard for CAC was defined in CSCT and CCTA. Out of these, 100 CCTA scans were used for training, and the remaining 100 for evaluation of a voxel classification method for CAC identification. The method uses ConvPairs, pairs of convolutional neural networks (ConvNets). The first ConvNet in a pair identifies voxels likely to be CAC, thereby discarding the majority of non-CAC-like voxels such as lung and fatty tissue. The identified CAC-like voxels are further classified by the second ConvNet in the pair, which distinguishes between CAC and CAC-like negatives. Given the different task of each ConvNet, they share their architecture, but not their weights. Input patches are either 2.5D or 3D. The ConvNets are purely convolutional, i.e. no pooling layers are present and fully connected layers are implemented as convolutions, thereby allowing efficient voxel classification. The performance of individual 2.5D and 3D ConvPairs with input sizes of 15 and 25 voxels, as well as the performance of ensembles of these ConvPairs, were evaluated by a comparison with reference annotations in CCTA and CSCT. In all cases, ensembles of ConvPairs outperformed their individual members. The best performing individual ConvPair detected 72% of lesions in the test set, with on average 0.85 false positive (FP) errors per scan. The best performing ensemble combined all ConvPairs and obtained a sensitivity of 71% at 0.48 FP errors per scan. For this ensemble, agreement with the reference mass score in CSCT was excellent (ICC 0.944 [0.918-0.962]). Aditionally, based on the Agatston score in CCTA, this ensemble assigned 83% of patients to the same cardiovascular risk category as reference CSCT. In conclusion, CAC can be accurately automatically identified and quantified in CCTA using the proposed pattern recognition method. This might obviate the need to acquire a dedicated CSCT scan for CAC scoring, which is regularly acquired prior to a CCTA, and thus reduce the CT radiation dose received by patients.

Automated assessment of thigh composition using machine learning for Dixon magnetic resonance images.

To develop and validate a machine learning based automated segmentation method that jointly analyzes the four contrasts provided by Dixon MRI technique for improved thigh composition segmentation accuracy. The automatic detection of body composition is formulized as a three-class classification issue. Each image voxel in the training dataset is assigned with a correct label. A voxel classifier is trained and subsequently used to predict unseen data. Morphological operations are finally applied to generate volumetric segmented images for different structures. We applied this algorithm on datasets of (1) four contrast images, (2) water and fat images, and (3) unsuppressed images acquired from 190 subjects. The proposed method using four contrasts achieved most accurate and robust segmentation compared to the use of combined fat and water images and the use of unsuppressed image, average Dice coefficients of 0.94±0.03, 0.96±0.03, 0.80±0.03, and 0.97±0.01 has been achieved to bone region, subcutaneous adipose tissue (SAT), inter-muscular adipose tissue (IMAT), and muscle respectively. Our proposed method based on machine learning produces accurate tissue quantification and showed an effective use of large information provided by the four contrast images from Dixon MRI.

Voxel-Based Diagnosis of Alzheimer's Disease Using Classifier Ensembles.

Functional magnetic resonance imaging (fMRI) is one of the most promising noninvasive techniques for early Alzheimer's disease (AD) diagnosis. In this paper, we explore the application of different machine learning techniques to the classification of fMRI data for this purpose. The functional images were first preprocessed using the statistical parametric mapping toolbox to output individual maps of statistically activated voxels. A fast filter was applied afterwards to select voxels commonly activated across demented and nondemented groups. Four feature ranking selection techniques were embedded into a wrapper scheme using an inner-outer loop for the selection of relevant voxels. The wrapper approach was guided by the performance of six pattern recognition models, three of which were ensemble classifiers based on stochastic searches. Final classification performance was assessed from the nested internal and external cross-validation loops taking several voxel sets ordered by importance. Numerical performance was evaluated using statistical tests, and the best combination of voxel selection and classification reached a 97.14% average accuracy. Results repeatedly pointed out Brodmann regions with distinct activation patterns between demented and nondemented profiles, indicating that the machine learning analysis described is a powerful method to detect differences in several brain regions between both groups.

Design and implementation of coded aperture coherent scatter spectral imaging of cancerous and healthy breast tissue samples.

A scatter imaging technique for the differentiation of cancerous and healthy breast tissue in a heterogeneous sample is introduced in this work. Such a technique has potential utility in intraoperative margin assessment during lumpectomy procedures. In this work, we investigate the feasibility of the imaging method for tumor classification using Monte Carlo simulations and physical experiments. The coded aperture coherent scatter spectral imaging technique was used to reconstruct three-dimensional (3-D) images of breast tissue samples acquired through a single-position snapshot acquisition, without rotation as is required in coherent scatter computed tomography. We perform a quantitative assessment of the accuracy of the cancerous voxel classification using Monte Carlo simulations of the imaging system; describe our experimental implementation of coded aperture scatter imaging; show the reconstructed images of the breast tissue samples; and present segmentations of the 3-D images in order to identify the cancerous and healthy tissue in the samples. From the Monte Carlo simulations, we find that coded aperture scatter imaging is able to reconstruct images of the samples and identify the distribution of cancerous and healthy tissues (i.e., fibroglandular, adipose, or a mix of the two) inside them with a cancerous voxel identification sensitivity, specificity, and accuracy of 92.4%, 91.9%, and 92.0%, respectively. From the experimental results, we find that the technique is able to identify cancerous and healthy tissue samples and reconstruct differential coherent scatter cross sections that are highly correlated with those measured by other groups using x-ray diffraction. Coded aperture scatter imaging has the potential to provide scatter images that automatically differentiate cancerous and healthy tissue inside samples within a time on the order of a minute per slice.

Automated detection of cerebral microbleeds in patients with Traumatic Brain Injury.

In this paper a Computer Aided Detection (CAD) system is presented to automatically detect Cerebral Microbleeds (CMBs) in patients with Traumatic Brain Injury (TBI). It is believed that the presence of CMBs has clinical prognostic value in TBI patients. To study the contribution of CMBs in patient outcome, accurate detection of CMBs is required. Manual detection of CMBs in TBI patients is a time consuming task that is prone to errors, because CMBs are easily overlooked and are difficult to distinguish from blood vessels.This study included 33 TBI patients. Because of the laborious nature of manually annotating CMBs, only one trained expert manually annotated the CMBs in all 33 patients. A subset of ten TBI patients was annotated by six experts. Our CAD system makes use of both Susceptibility Weighted Imaging (SWI) and T1 weighted magnetic resonance images to detect CMBs. After pre-processing these images, a two-step approach was used for automated detection of CMBs. In the first step, each voxel was characterized by twelve features based on the dark and spherical nature of CMBs and a random forest classifier was used to identify CMB candidate locations. In the second step, segmentations were made from each identified candidate location. Subsequently an object-based classifier was used to remove false positive detections of the voxel classifier, by considering seven object-based features that discriminate between spherical objects (CMBs) and elongated objects (blood vessels). A guided user interface was designed for fast evaluation of the CAD system result. During this process, an expert checked each CMB detected by the CAD system.A Fleiss' kappa value of only 0.24 showed that the inter-observer variability for the TBI patients in this study was very large. An expert using the guided user interface reached an average sensitivity of 93%, which was significantly higher (p = 0.03) than the average sensitivity of 77% (sd 12.4%) that the six experts manually detected. Furthermore, with the use of this CAD system the reading time was substantially reduced from one hour to 13 minutes per patient, because the CAD system only detects on average 25.9 false positives per TBI patient, resulting in 0.29 false positives per definite CMB finding.

Redistribution of inhaled hyperpolarized 3He gas during breath-hold differs by asthma severity.

The purpose of this work was to quantify the redistribution of ventilation-weighted signal in the lungs of asthmatic subjects during a breath-hold using high temporal-spatial resolution hyperpolarized (HP) He-3 MRI. HP He-3 MRI was used to obtain time-resolved, volumetric images of lung ventilation during breath-hold in 39 human subjects classified as either healthy/nondiseased (n = 14), mild-to-moderate asthmatic (n = 17), or severely asthmatic (n = 8). Signals were normalized to a standard lung volume, so that voxels within the lung from all 39 subjects could be analyzed as a group to increase statistical power and enable semiautomated classification of voxels into 1 of 5 ventilation level categories (ranging from defect to hyperintense). End-inspiratory ventilation distribution and temporal rates of mean signal change for each of the five ventilation categories were compared using ANOVA. Time rates of signal change were hypothesized to represent underlying gas redistribution processes, potentially influenced by disease. We found that mild-to-moderate asthmatic subjects showed the greatest rate of signal change, even though those with severe asthma had the greatest end-inspiration ventilation heterogeneity. The observed results support the existence of local differences in airway resistances associated with the different obstructive patterns in the lungs for severe vs. mild-to-moderate asthmatic subjects.

MRI-Only Based Radiotherapy Treatment Planning for the Rat Brain on a Small Animal Radiation Research Platform (SARRP).

Computed tomography (CT) is the standard imaging modality in radiation therapy treatment planning (RTP). However, magnetic resonance (MR) imaging provides superior soft tissue contrast, increasing the precision of target volume selection. We present MR-only based RTP for a rat brain on a small animal radiation research platform (SARRP) using probabilistic voxel classification with multiple MR sequences. Six rat heads were imaged, each with one CT and five MR sequences. The MR sequences were: T1-weighted, T2-weighted, zero-echo time (ZTE), and two ultra-short echo time sequences with 20 μs (UTE1) and 2 ms (UTE2) echo times. CT data were manually segmented into air, soft tissue, and bone to obtain the RTP reference. Bias field corrected MR images were automatically segmented into the same tissue classes using a fuzzy c-means segmentation algorithm with multiple images as input. Similarities between segmented CT and automatic segmented MR (ASMR) images were evaluated using Dice coefficient. Three ASMR images with high similarity index were used for further RTP. Three beam arrangements were investigated. Dose distributions were compared by analysing dose volume histograms. The highest Dice coefficients were obtained for the ZTE-UTE2 combination and for the T1-UTE1-T2 combination when ZTE was unavailable. Both combinations, along with UTE1-UTE2, often used to generate ASMR images, were used for further RTP. Using 1 beam, MR based RTP underestimated the dose to be delivered to the target (range: 1.4%-7.6%). When more complex beam configurations were used, the calculated dose using the ZTE-UTE2 combination was the most accurate, with 0.7% deviation from CT, compared to 0.8% for T1-UTE1-T2 and 1.7% for UTE1-UTE2. The presented MR-only based workflow for RTP on a SARRP enables both accurate organ delineation and dose calculations using multiple MR sequences. This method can be useful in longitudinal studies where CT’s cumulative radiation dose might contribute to the total dose.

Potential for Early Fracture Risk Assessment in Patients with Metastatic Bone Disease using Parametric Response Mapping of CT Images.

Pathologic vertebral compression fractures (PVCFs) cause significant morbidity in patients with metastatic bone disease. Limitations in existing clinical biomarkers leave clinicians without reliable metrics for predicting PVCF, thus impeding efforts to prevent this severe complication. To establish the feasibility of a new method for defining the risk of a PVCF, we retrospectively analyzed serial computed tomography (CT) scans from 5 breast cancer patients using parametric response mapping (PRM) to quantify dynamic bone miniral density (BMD) changes that preceded an event. Vertebrae segmented from each scan were registered to the same spatial frame and voxel classification was accomplished using a predetermined threshold of change in Hounsfield units (HU), resulting in relative volumes of increased (PRMHU+), decreased (PRMHU−), or unchanged (PRMHU0) attenuation. A total of 7 PVCFs were compared to undiseased vertebrae in each patient serving as controls. A receiver operator curve (ROC) analysis identified optimal imaging times for group stratification. BMD changes were apparent by an elevated PRMHU+ as early as 1 year before fracture. ROC analysis showed poor performance of PRMHU− in stratifying PVCFs versus controls. As early as 6 months before PVCF, PRMHU+ was significantly larger (12.9 ± 11.6%) than control vertebrae (2.3 ± 2.5%), with an area under the curve of 0.918 from an ROC analysis. Mean HU changes were also significant between PVCF (26.8 ± 26.9%) and control (−2.2 ± 22.0%) over the same period. A PRM analysis of BMD changes using standard CT imaging was sensitive for spatially resolving changes that preceded structural failure in these patients.

A multiparametric and multiscale approach to automated segmentation of brain veins.

Cerebral vein analysis provides a fundamental tool to study brain diseases such as neurodegenerative disorders or traumatic brain injuries. In order to assess the vascular anatomy, manual segmentation approaches can be used but are observer-dependent and time-consuming. In the present work, a fully automated cerebral vein segmentation method is proposed, based on a multiscale and multiparametric approach. The combined investigation of the R2(*)- and a Vesselness probability-map was used to obtain a fast and highly reliable classification of venous voxels. A semiquantitative analysis showed that our approach outperformed the previous state-of-the-art algorithm both in sensitivity and specificity. Inclusion of this tool within a parametric brain framework may therefore pave the way for a quantitative study of the intracranial venous system.

Automatic segmentation of MR brain images of preterm infants using supervised classification

Preterm birth is often associated with impaired brain development. The state and expected progression of preterm brain development can be evaluated using quantitative assessment of MR images. Such measurements require accurate segmentation of different tissue types in those images.This paper presents an algorithm for the automatic segmentation of unmyelinated white matter (WM), cortical grey matter (GM), and cerebrospinal fluid in the extracerebral space (CSF). The algorithm uses supervised voxel classification in three subsequent stages. In the first stage, voxels that can easily be assigned to one of the three tissue types are labelled. In the second stage, dedicated analysis of the remaining voxels is performed. The first and the second stages both use two-class classification for each tissue type separately. Possible inconsistencies that could result from these tissue-specific segmentation stages are resolved in the third stage, which performs multi-class classification. A set of T1- and T2-weighted images was analysed, but the optimised system performs automatic segmentation using a T2-weighted image only.We have investigated the performance of the algorithm when using training data randomly selected from completely annotated images as well as when using training data from only partially annotated images. The method was evaluated on images of preterm infants acquired at 30 and 40weeks postmenstrual age (PMA). When the method was trained using random selection from the completely annotated images, the average Dice coefficients were 0.95 for WM, 0.81 for GM, and 0.89 for CSF on an independent set of images acquired at 30weeks PMA. When the method was trained using only the partially annotated images, the average Dice coefficients were 0.95 for WM, 0.78 for GM and 0.87 for CSF for the images acquired at 30weeks PMA, and 0.92 for WM, 0.80 for GM and 0.85 for CSF for the images acquired at 40weeks PMA. Even though the segmentations obtained using training data from the partially annotated images resulted in slightly lower Dice coefficients, the performance in all experiments was close to that of a second human expert (0.93 for WM, 0.79 for GM and 0.86 for CSF for the images acquired at 30weeks, and 0.94 for WM, 0.76 for GM and 0.87 for CSF for the images acquired at 40weeks).These results show that the presented method is robust to age and acquisition protocol and that it performs accurate segmentation of WM, GM, and CSF when the training data is extracted from complete annotations as well as when the training data is extracted from partial annotations only. This extends the applicability of the method by reducing the time and effort necessary to create training data in a population with different characteristics.

Pseudo-CT generation in brain MR-PET attenuation correction: comparison of several multi-atlas methods.

Simultaneous MR-PET imaging opens new perspectives for understanding many aspects of brain function. However, accurate brain attenuation correction (AC) is required for absolute PET quantification. Radiodensity maps for AC are not readily available in PET-MR. Various strategies have been proposed to generate substitute attenuation maps. Recently, it has been shown that multiatlas techniques for AC outperform methods based on MR imaging sequences, such as Ultrashort-Echo-Time. We recently investigated several multiatlas methods to build a patient-specific pseudoCT, all based on multiatlas registration and label propagation, and introduced a novel maximum probability method (MaxProb). Here, a new database of 40 MR/CT image pairs (atlases) was used. Pairwise nonrigid registration of each atlas MR image to the target MR image yielded a geometric transformation that we used for propagating the atlas CT into the target’s space. The synthetic CT was generated by voxelwise atlas selection and intensity averaging. We implemented a state-of-the-art algorithm along with two variants and MaxProb. A pseudoCT was generated for each subject of the database and assessed by comparison with the real CT in a leave-one-out scheme. Jaccard overlap indices were computed per tissue class (air, soft tissue, and bone). Voxel classification error was also assessed. The Jaccard indices and voxel classification error showed that MaxProb (8.31±1.12%) performed slightly better than Ref (8.65±0.93%). The influence of the number of atlases fused in pseudoCT generation was studied, showing that an optimal number improves accuracy. By adjusting the number of fused atlases, it was possible to significantly improve on Ref (8.46±0.96%). The methods now need evaluation on quantitative PET data to identify potential benefits in PET analysis.

Automatic segmentation of high- and low-field knee MRIs using knee image quantification with data from the osteoarthritis initiative.

Clinical studies including thousands of magnetic resonance imaging (MRI) scans offer potential for pathogenesis research in osteoarthritis. However, comprehensive quantification of all bone, cartilage, and meniscus compartments is challenging. We propose a segmentation framework for fully automatic segmentation of knee MRI. The framework combines multiatlas rigid registration with voxel classification and was trained on manual segmentations with varying configurations of bones, cartilages, and menisci. The validation included high- and low-field knee MRI cohorts from the Center for Clinical and Basic Research, the osteoarthritis initiative (QAI), and the segmentation of knee images10 (SKI10) challenge. In total, 1907 knee MRIs were segmented during the evaluation. No segmentations were excluded. Our resulting OAI cartilage volume scores are available upon request. The precision and accuracy performances matched manual reader re-segmentation well. The cartilage volume scan-rescan precision was 4.9% (RMS CV). The Dice volume overlaps in the medial/lateral tibial/femoral cartilage compartments were 0.80 to 0.87. The correlations with volumes from independent methods were between 0.90 and 0.96 on the OAI scans. Thus, the framework demonstrated precision and accuracy comparable to manual segmentations. Finally, our method placed second for cartilage segmentation in the SKI10 challenge. The comprehensive validation suggested that automatic segmentation is appropriate for cohorts with thousands of scans.

Stratified Sampling Voxel Classification for Segmentation of Intraretinal and Subretinal Fluid in Longitudinal Clinical OCT Data.

Automated three-dimensional retinal fluid (named symptomatic exudate-associated derangements, SEAD) segmentation in 3D OCT volumes is of high interest in the improved management of neovascular Age Related Macular Degeneration (AMD). SEAD segmentation plays an important role in the treatment of neovascular AMD, but accurate segmentation is challenging because of the large diversity of SEAD size, location, and shape. Here a novel voxel classification based approach using a layer-dependent stratified sampling strategy was developed to address the class imbalance problem in SEAD detection. The method was validated on a set of 30 longitudinal 3D OCT scans from 10 patients who underwent anti-VEGF treatment. Two retinal specialists manually delineated all intraretinal and subretinal fluid. Leave-one-patient-out evaluation resulted in a true positive rate and true negative rate of 96% and 0.16% respectively. This method showed promise for image guided therapy of neovascular AMD treatment.

Pseudo-CT generation from multiple MR images for small animal irradiation

Introduction. Computed tomography (CT) is the standard imaging modality for radiation therapy treatment planning (RTTP) because of its ability to provide information on electron density. However, magnetic resonance (MR) imaging provides superior soft tissue contrast, especially in small animal imaging, facilitating the precise selection of the target volume. This makes the technique interesting for irradiation of brain tumors. The aim of this study was to present an MR-only based workflow for RTTP on a small animal radiation research platform (SARRP) by investigating the potential of probabilistic classification of voxels using multiple MR sequences. Methods. Six female Fisher rats were anesthetized using isoflurane and individually fixed on an in-house made multimodality bed before starting MR and CT acquisitions. MR measurements were performed on a 7-Tesla system using a rat brain volume coil. Four different MR sequences were acquired for each animal, including a T1-weighted (MDEFT) sequence, a T2-weighted (RARE) sequence, an ultra-short echo time sequence with 20 μs echo time (UTE1) and an ultra- short echo time sequence with 2 ms echo time (UTE2). UTE offers the opportunity to acquire images from proton-poor structures with very short transverse relaxation times, such as bone, by using a rapid readout of the fast decaying signal. Following MR, the animals were moved to the SARRP to start a cone-beam CT (CB-CT) by acquiring 720 projections over 360°. Cone-beam CT projection data were reconstructed by filtered back-projection to obtain the standard-CT for RTTP. Then the images were bias field corrected and manually co-registered to the CB-CT. After that, images were segmented in three tissue classes (air, soft tissue and bone) with k-means for the CB-CT and fuzzy c-means segmentation algorithm (FCM) for the MR images with multiple MR images as input. The membership probability can be between 0 and 1, with one indicating 100% probability and zero indicating 0% probability to belong to a specific tissue class. To obtain a pseudo-CT image, voxels were assigned to the tissue class having the highest membership probability. The dice coefficient was used to evaluate the correctness of the segmentation for soft-tissue and bone. The pseudo-CT images with the highest similarity index were used for further radiotherapy treatment planning (RTTP), in addition, to the standard UTE1-UTE2. The target of the RTTP that was selected in the primary cortex (M1) and three different beam arrangements were investigated to compare CB-CT and MR-based dose calculations. The dose plans were a single static beam of 3x3 mm, using a single arc (3x3 mm beam size, 120° arc, couch at 0°), and three non-coplanar arcs (3x3 mm beam size, 120° arc, couch at 0°, 45° and 90°). Dose distributions were calculated using the TPS of the SARRP and cumulative dose volume histograms (DVHs) of the target and normal brain tissue were obtained for the three dose plans. Results The highest dice coefficient was obtained for the T1-UTE1-T2 combination, which was used for further RTTP. The contribution of bone to the total dice coefficient did not exceed 27%. However, bone accounts for only 2% of the image, therefore a misclassified bone pixel has a bigger effect in the dice coefficient than a misclassified soft tissue pixel. Using only 1 beam, both MR combinations underestimate the dose to be delivered to the target. When more complex beam configurations were used to irradiate the target, very small differences were observed between CB-CT and MR based dose calculations.

BOOST: a supervised approach for multiple sclerosis lesion segmentation.

Automatic multiple sclerosis lesion segmentation is a challenging task. An extensive analysis of the most recent techniques indicates an improvement of the results obtained when using prior knowledge and contextual information. We present BOOST, a knowledge-based approach to automatically segment multiple sclerosis lesions through a voxel by voxel classification. We used the Gentleboost classifier and a set of features, including contextual features, registered atlas probability maps and an outlier map. Results are computed on a set of 45 cases from three different hospitals (15 of each), obtaining a moderate agreement between the manual annotations and the automatically segmented results. We quantitatively compared our results with three public state-of-the-art approaches obtaining competitive results and a better overlap with manual annotations. Our approach tends to better segment those cases with high lesion load, while cases with small lesion load are more difficult to accurately segment. We believe BOOST has potential applicability in the clinical practice, although it should be improved in those cases with small lesion load.

Automated voxel-by-voxel tissue classification for hippocampal segmentation: Methods and validation

The hippocampus is an important structural biomarker for Alzheimer's disease (AD) and has a primary role in the pathogenesis of other neurological and psychiatric diseases. This study presents a fully automated pattern recognition system for an accurate and reproducible segmentation of the hippocampus in structural Magnetic Resonance Imaging (MRI). The method was validated on a mixed cohort of 56 T1-weighted structural brain images, and consists of three processing levels: (a) Linear registration: all brain images were registered to a standard template and an automated method was applied to capture the global shape of the hippocampus. (b) Feature extraction: all voxels included in the previously selected volume were characterized by 315 features computed from local information. (c) Voxel classification: a Random Forest algorithm was used to classify voxels as belonging or not belonging to the hippocampus. In order to improve the classification performance, an adaptive learning method based on the use of the Pearson's correlation coefficient was developed. The segmentation results (Dice similarity index = 0.81 ± 0.03) compare well with other state-of-the art approaches. A validation study was conducted on an independent dataset of 100 T1-weighted brain images, achieving significantly better results than those obtained with FreeSurfer.

Computer-aided detection of prostate cancer in MRI.

Prostate cancer is one of the major causes of cancer death for men in the western world. Magnetic resonance imaging (MRI) is being increasingly used as a modality to detect prostate cancer. Therefore, computer-aided detection of prostate cancer in MRI images has become an active area of research. In this paper we investigate a fully automated computer-aided detection system which consists of two stages. In the first stage, we detect initial candidates using multi-atlas-based prostate segmentation, voxel feature extraction, classification and local maxima detection. The second stage segments the candidate regions and using classification we obtain cancer likelihoods for each candidate. Features represent pharmacokinetic behavior, symmetry and appearance, among others. The system is evaluated on a large consecutive cohort of 347 patients with MR-guided biopsy as the reference standard. This set contained 165 patients with cancer and 182 patients without prostate cancer. Performance evaluation is based on lesion-based free-response receiver operating characteristic curve and patient-based receiver operating characteristic analysis. The system is also compared to the prospective clinical performance of radiologists. Results show a sensitivity of 0.42, 0.75, and 0.89 at 0.1, 1, and 10 false positives per normal case. In clinical workflow the system could potentially be used to improve the sensitivity of the radiologist. At the high specificity reading setting, which is typical in screening situations, the system does not perform significantly different from the radiologist and could be used as an independent second reader instead of a second radiologist. Furthermore, the system has potential in a first-reader setting.

A clonal selection based approach to statistical brain voxel classification in magnetic resonance images

Statistical classification of voxels in brain magnetic resonance (MR) images into major tissue types plays an important role in neuroscience research and clinical practices, in which model estimation is an essential step. Despite their prevalence, traditional techniques, such as the expectation–maximization (EM) algorithm and genetic algorithm (GA), have inherent limitations, and may result in less-accurate classification. In this paper, we introduce the immune-inspired clonal selection algorithm (CSA) to the maximum likelihood estimation of the Gaussian mixture model (GMM), and thus propose the GMM-CSA algorithm for automated voxel classification in brain MR images. This algorithm achieves simultaneous voxel classification and bias field correction in a three-stage iterative process under the CSA framework. At each iteration, a population of admissible model parameters, voxel labels and estimated bias field are updated. To explore the prior anatomical knowledge, we also construct a probabilistic brain atlas for each MR study and incorporate the atlas into the classification process. The GMM-CSA algorithm has been compared to five state-of-the-art brain MR image segmentation approaches on both simulated and clinical data. Our results show that the proposed algorithm is capable of classifying voxels in brain MR images into major tissue types more accurately.

Anatomic segmentation improves prostate cancer detection with artificial neural networks analysis of1H magnetic resonance spectroscopic imaging

PurposeTo assess whether an artificial neural network (ANN) model is a useful tool for automatic detection of cancerous voxels in the prostate from1H‐MRSI datasets and whether the addition of information about anatomical segmentation improves the detection of cancer.Materials and MethodsThe Institutional Review Board approved this HIPAA‐compliant study and waived informed consent. Eighteen men with prostate cancer (median age, 55 years; range, 36–71 years) who underwent endorectal MRI/MRSI before radical prostatectomy were included in this study. These patients had at least one cancer area on whole—mount histopathological map and at least one matching MRSI voxel suspicious for cancer detected. Two ANN models for automatic classification of MRSI voxels in the prostate were implemented and compared: model 1, which used only spectra as input, and model 2, which used the spectra plus information from anatomical segmentation. The models were trained, tested and validated using spectra from voxels that the spectroscopist had designated as cancer and that were verified on histopathological maps.ResultsAt ROC analysis, model 2 (AUC = 0.968) provided significantly better (P= 0.03) classification of cancerous voxels than did model 1 (AUC = 0.949).ConclusionAutomatic analysis of prostate MRSI to detect cancer using ANN model is feasible. Application of anatomical segmentation from MRI as an additional input to ANN improves the accuracy of detecting cancerous voxels from MRSI.J. Magn. Reson. Imaging 2014;40:1414–1421. ©2013 Wiley Periodicals, Inc.

Investigation of a method for generating synthetic CT models from MRI scans of the head and neck for radiation therapy

Magnetic resonance (MR) images often provide superior anatomic and functional information over computed tomography (CT) images, but generally are not used alone without CT images for radiotherapy treatment planning and image guidance. This study aims to investigate the potential of probabilistic classification of voxels from multiple MRI contrasts to generate synthetic CT (‘MRCT’) images. The method consists of (1) acquiring multiple MRI volumes: T1-weighted, T2-weighted, two echoes from a ultra-short echo time (UTE) sequence, and calculated fat and water image volumes using a Dixon method, (2) classifying tissues using fuzzy c-means clustering with a spatial constraint, (3) assigning attenuation properties with weights based on the probability of individual tissue classes being present in each voxel, and (4) generating a MRCT image volume from the sum of attenuation properties in each voxel. The capability of each MRI contrast to differentiate tissues of interest was investigated based on a retrospective analysis of ten patients. For one prospective patient, the correlation of skull intensities between CT and MR was investigated, the discriminatory power of MRI in separating air from bone was evaluated, and the generated MRCT image volume was qualitatively evaluated. Our analyses showed that one MRI volume was not sufficient to separate all tissue types, and T2-weighted images was more sensitive to bone density variation compared to other MRI image types. The short echo UTE image showed significant improvement in contrasting air versus bone, but could not completely separate air from bone without false labeling. Generated MRCT and CT images showed similar contrast between bone and soft/solid tissues. These results demonstrate the potential of the presented method to generate synthetic CT images to support the workflow of radiation oncology treatment planning and image guidance.