Automatic quality control of brain T1-weighted magnetic resonance images for a clinical data warehouse

Many studies on machine learning (ML) for computer-aided diagnosis have so far been mostly restricted to high-quality research data. Clinical data warehouses, gathering routine examinations from hospitals, offer great promises for training and validation of ML models in a realistic setting. However, the use of such clinical data warehouses requires quality control (QC) tools. Visual QC by experts is time-consuming and does not scale to large datasets. In this paper, we propose a convolutional neural network (CNN) for the automatic QC of 3D T1-weighted brain MRI for a large heterogeneous clinical data warehouse. To that purpose, we used the data warehouse of the hospitals of the Greater Paris area (Assistance Publique-Hôpitaux de Paris [AP-HP]). Specifically, the objectives were: 1) to identify images which are not proper T1-weighted brain MRIs; 2) to identify acquisitions for which gadolinium was injected; 3) to rate the overall image quality. We used 5000 images for training and validation and a separate set of 500 images for testing. In order to train/validate the CNN, the data were annotated by two trained raters according to a visual QC protocol that we specifically designed for application in the setting of a data warehouse. For objectives 1 and 2, our approach achieved excellent accuracy (balanced accuracy and F1-score >90%), similar to the human raters. For objective 3, the performance was good but substantially lower than that of human raters. Nevertheless, the automatic approach accurately identified (balanced accuracy and F1-score >80%) low quality images, which would typically need to be excluded. Overall, our approach shall be useful for exploiting hospital data warehouses in medical image computing.     

Semi-automatic segmentation of myocardium at risk in T2-weighted cardiovascular magnetic resonance

Background

T2-weighted cardiovascular magnetic resonance (CMR) has been shown to be a promising technique for determination of ischemic myocardium, referred to as myocardium at risk (MaR), after an acute coronary event. Quantification of MaR in T2-weighted CMR has been proposed to be performed by manual delineation or the threshold methods of two standard deviations from remote (2SD), full width half maximum intensity (FWHM) or Otsu. However, manual delineation is subjective and threshold methods have inherent limitations related to threshold definition and lack of a priori information about cardiac anatomy and physiology. Therefore, the aim of this study was to develop an automatic segmentation algorithm for quantification of MaR using anatomical a priori information.

Methods

Forty-seven patients with first-time acute ST-elevation myocardial infarction underwent T2-weighted CMR within 1 week after admission. Endocardial and epicardial borders of the left ventricle, as well as the hyper enhanced MaR regions were manually delineated by experienced observers and used as reference method. A new automatic segmentation algorithm, called Segment MaR, defines the MaR region as the continuous region most probable of being MaR, by estimating the intensities of normal myocardium and MaR with an expectation maximization algorithm and restricting the MaR region by an a priori model of the maximal extent for the user defined culprit artery. The segmentation by Segment MaR was compared against inter observer variability of manual delineation and the threshold methods of 2SD, FWHM and Otsu.

Results

MaR was 32.9 ± 10.9% of left ventricular mass (LVM) when assessed by the reference observer and 31.0 ± 8.8% of LVM assessed by Segment MaR. The bias and correlation was, -1.9 ± 6.4% of LVM, R = 0.81 (p < 0.001) for Segment MaR, -2.3 ± 4.9%, R = 0.91 (p < 0.001) for inter observer variability of manual delineation, -7.7 ± 11.4%, R = 0.38 (p = 0.008) for 2SD, -21.0 ± 9.9%, R = 0.41 (p = 0.004) for FWHM, and 5.3 ± 9.6%, R = 0.47 (p < 0.001) for Otsu.

Conclusions

There is a good agreement between automatic Segment MaR and manually assessed MaR in T2-weighted CMR. Thus, the proposed algorithm seems to be a promising, objective method for standardized MaR quantification in T2-weighted CMR.     

Fast and robust multi-atlas segmentation of brain magnetic resonance images

We introduce an optimised pipeline for multi-atlas brain MRI segmentation. Both accuracy and speed of segmentation are considered. We study different similarity measures used in non-rigid registration. We show that intensity differences for intensity normalised images can be used instead of standard normalised mutual information in registration without compromising the accuracy but leading to threefold decrease in the computation time. We study and validate also different methods for atlas selection. Finally, we propose two new approaches for combining multi-atlas segmentation and intensity modelling based on segmentation using expectation maximisation (EM) and optimisation via graph cuts. The segmentation pipeline is evaluated with two data cohorts: IBSR data (N = 18, six subcortial structures: thalamus, caudate, putamen, pallidum, hippocampus, amygdala) and ADNI data (N =   60, hippocampus). The average similarity index between automatically and manually generated volumes was 0.849 (IBSR, six subcortical structures) and 0.880 (ADNI, hippocampus). The correlation coefficient for hippocampal volumes was 0.95 with the ADNI data. The computation time using a standard multicore PC computer was about 3–4 min. Our results compare favourably with other recently published results.     

Sequence-independent segmentation of magnetic resonance images

We present a set of techniques for embedding the physics of the imaging process that generates a class of magnetic resonance images (MRIs) into a segmentation or registration algorithm. This results in substantial invariance to acquisition parameters, as the effect of these parameters on the contrast properties of various brain structures is explicitly modeled in the segmentation. In addition, the integration of image acquisition with tissue classification allows the derivation of sequences that are optimal for segmentation purposes. Another benefit of these procedures is the generation of probabilistic models of the intrinsic tissue parameters that cause MR contrast (e.g., T1, proton density, T2*), allowing access to these physiologically relevant parameters that may change with disease or demographic, resulting in nonmorphometric alterations in MR images that are otherwise difficult to detect. Finally, we also present a high band width multiecho FLASH pulse sequence that results in high signal-to-noise ratio with minimal image distortion due to B0 effects. This sequence has the added benefit of allowing the explicit estimation of T2* and of reducing test-retest intensity variability.     

Phantom-based performance evaluation: application to brain segmentation from magnetic resonance images

This paper presents a new technique for assessing the accuracy of segmentation algorithms, applied to the performance evaluation of brain editing and brain tissue segmentation algorithms for magnetic resonance images. We propose performance evaluation criteria derived from the use of the realistic digital brain phantom Brainweb. This 'ground truth' allows us to build distance-based discrepancy features between the edited brain or the segmented brain tissues (such as cerebro-spinal fluid, grey matter and white matter) and the phantom model, taken as a reference. Furthermore, segmentation errors can be spatially determined, and ranged in terms of their distance to the reference. The brain editing method used is the combination of two segmentation techniques. The first is based on binary mathematical morphology and a region growing approach. It represents the initialization step, the results of which are then refined with the second method, using an active contour model. The brain tissue segmentation used is based on a Markov random field model. Segmentation results are shown on the phantom for each method, and on real magnetic resonance images for the editing step; performance is evaluated by the new distance-based technique and corroborates the effective refinement of the segmentation using active contours. The criteria described here can supersede biased visual inspection in order to compare, evaluate and validate any segmentation algorithm. Moreover, provided a 'ground truth' is given, we are able to determine quantitatively to what extent a segmentation algorithm is sensitive to internal parameters, noise, artefacts or distortions.     

Validation of partial tissue segmentation of single-channel magnetic resonance images of the brain

We describe and evaluate a practical, automated algorithm based on local statistical mixture modeling for segmenting single-channel, T1-weighted volumetric magnetic resonance images of the brain into gray matter, white matter, and cerebrospinal fluid. We employed a stereological sampling method to assess, prospectively, the performance of the method with respect to human experts on 10 normal T1-weighted brain scans acquired with a three-dimensional gradient echo pulse sequence. The overall kappa statistic for the concordance of the algorithm with the human experts was 0.806, while that among raters, excluding the algorithm, was 0.802. The algorithm had better agreement with the modal expert decision (kappa = 0.878). The algorithm could not be distinguished from the experts by this measure. We also validated the algorithm on a simulated MR scan of a digital brain phantom with known tissue composition. Global gray matter and white matter errors were 1% and <1%, respectively, and correlation coefficients with the underlying tissue model were 0.95 for gray matter, 0.98 for white matter, and 0.95 for cerebrospinal fluid. In both approaches to validation, we evaluated both local and global performance of the algorithm. Human experts generated slightly higher global gray matter proportion estimates on the test brain scans relative to the algorithm (3.7%) and on the simulated MR scan relative to the true tissue model (4.4%). The algorithm underestimated gray in some subcortical nuclei which contain admixed gray and white matter. We demonstrate the reliability of the method on individual 1 NEX data sets of the test subjects, and its insensitivity to the precise values of initial model parameters. The output of this algorithm is suitable for quantifying cerebral cortical tissue, using a commonly performed commercial pulse sequence.     

T1 and T2 relaxation times on gadolinium-diethylenetriaminepentaacetic acid enhanced magnetic resonance images of brain tumors

MRI of brain tumors was performed with CPMG pulse sequence, and T1 and T2 relaxation times were measured before and after the administration of 0.1 mmol/kg Gd-DTPA. T1 and T2 became significantly shorter than before with the T1 shortening being about twice the T2's. Three types of T2 shortening were classified by the difference in the intra-tumoral transport of Gd-DTPA in the time-course study after the administration. Transportal activity of Gd-DTPA seemed to be reflected by the physiologic viability of tumor suggested by pre-contrast T2.     

T2-weighted cardiovascular magnetic resonance in acute cardiac disease

Cardiovascular magnetic resonance (CMR) using T2-weighted sequences can visualize myocardial edema. When compared to previous protocols, newer pulse sequences with substantially improved image quality have increased its clinical utility. The assessment of myocardial edema provides useful incremental diagnostic and prognostic information in a variety of clinical settings associated with acute myocardial injury. In patients with acute chest pain, T2-weighted CMR is able to identify acute or recent myocardial ischemic injury and has been employed to distinguish acute coronary syndrome (ACS) from non-ACS as well as acute from chronic myocardial infarction.T2-weighted CMR can also be used to determine the area at risk in reperfused and non-reperfused infarction. When combined with contrast-enhanced imaging, the salvaged area and thus the success of early coronary revascularization can be quantified. Strong evidence for the prognostic value of myocardial salvage has enabled its use as a primary endpoint in clinical trials. The present article reviews the current evidence and clinical applications for T2-weighted CMR in acute cardiac disease and gives an outlook on future developments."The principle of all things is water"Thales of Miletus (624 BC - 546 BC)     

Visualizing human brain surface from T1-weighted MR images using texture-mapped triangle meshes

We describe a novel method for visualizing brain surface from anatomical magnetic resonance images (MRIs). The method utilizes standard 2D texture mapping capabilities of OpenGL graphics language. It combines the benefits of volume rendering and triangle-mesh rendering, allowing fast and realistic-looking brain surface visualizations. Consequently, relatively low-resolution triangle meshes can be used while the texture images provide the necessary details. The mapping is optimized to provide good texture-image resolution for the triangles with respect to their original sizes in the 3D MRI volume. The actual 2D texture images are generated by depth integration from the original MRI data. Our method adapts to anisotropic voxel sizes without any need to interpolate the volume data into cubic voxels, and it is very well suited for visualizing brain anatomy from standard T(1)-weighted MR images. Furthermore, other OpenGL objects and techniques can be easily combined, for example, to use cut planes, to show other surfaces and objects, and to visualize functional data in addition to the anatomical information.     

Semiautomatic segmentation of brain exterior in magnetic resonance images driven by empirical procedures and anatomical knowledge

This work demonstrates encouraging results for increasing the automation of a practical and precise magnetic resonance brain image segmentation method. The intensity threshold for segmenting the brain exterior is determined automatically by locating the choroid plexus. This is done by finding peaks in a series of histograms taken over regions specified using anatomical knowledge. Intensity inhomogeneities are accounted for by adjusting the global intensity to match the white matter peak intensity in local regions. Automated results are incorporated into the established manually guided segmentation method by providing a trained expert with the automated threshold. The results from 20 different brain scans (over 1000 images) obtained under different conditions are presented to validate the method which was able to determine the appropriate threshold in approximately 80% of the data.     

Segmentation of brain tissue from magnetic resonance images

Segmentation of medical imagery is a challenging problem due to the complexity of the images, as well as to the absence of models of the anatomy that fully capture the possible deformations in each structure. The brain is a particularly complex structure, and its segmentation is an important step for many problems, including studies in temporal change detection of morphology, and 3-D visualizations for surgical planning. We present a method for segmentation of brain tissue from magnetic resonance images that is a combination of three existing techniques from the computer vision literature: expectation/maximization segmentation, binary mathematical morphology, and active contour models. Each of these techniques has been customized for the problem of brain tissue segmentation such that the resultant method is more robust than its components. Finally, we present the results of a parallel implementation of this method on IBM's supercomputer Power Visualization System for a database of 20 brain scans each with 256 × 256 × 124 voxels and validate those results against segmentations generated by neuroanatomy experts.     

Automated left ventricular segmentation from cardiac magnetic resonance images via adversarial learning with multi-stage pose estimation network and co-discriminator

Left ventricular (LV) segmentation is essential for the early diagnosis of cardiovascular diseases, which has been reported as the leading cause of death all over the world. However, automated LV segmentation from cardiac magnetic resonance images (CMRI) using the traditional convolutional neural networks (CNNs) is still a challenging task due to the limited labeled CMRI data and low tolerances to irregular scales, shapes and deformations of LV. In this paper, we propose an automated LV segmentation method based on adversarial learning by integrating a multi-stage pose estimation network (MSPN) and a co-discrimination network. Different from existing CNNs, we use a MSPN with multi-scale dilated convolution (MDC) modules to enhance the ranges of receptive field for deep feature extraction. To fully utilize both labeled and unlabeled CMRI data, we propose a novel generative adversarial network (GAN) framework for LV segmentation by combining MSPN with co-discrimination networks. Specifically, the labeled CMRI are first used to initialize our segmentation network (MSPN) and co-discrimination network. Our GAN training includes two different kinds of epochs fed with both labeled and unlabeled CMRI data alternatively, which are different from the traditional CNNs only relied on the limited labeled samples to train the segmentation networks. As both ground truth and unlabeled samples are involved in guiding training, our method not only can converge faster but also obtain a better performance in LV segmentation. Our method is evaluated using MICCAI 2009 and 2017 challenge databases. Experimental results show that our method has obtained promising performance in LV segmentation, which also outperforms the state-of-the-art methods in terms of LV segmentation accuracy from the comparison results.     

Assessment of intramyocardial hemorrhage by T1-weighted cardiovascular magnetic resonance in reperfused acute myocardial infarction

Background

Intramyocardialhemorrhage (IMH) reflects severe reperfusion injury in acute myocardial infarction. Non-invasive detection of IMH by cardiovascular magnetic resonance (CMR) may serve as a surrogate marker to evaluate the effect of preventive measures to reduce reperfusion injury and hence provide additional prognostic information. We sought to investigate whether IMH could be detected by CMR exploiting the T1 shortening effect of methemoglobin in an experimental model of acute myocardial infarction. The results were compared to T2-weighthed short tau inversion recovery (T2-STIR), and T2*-weighted(T2*W) sequences.

Methods and results

IMH was induced in ten 40 kg pigs by 50-min balloon occlusion of the mid LAD followed by reperfusion. Between 4–9 days (average 4.8) post-injury, the left ventricular myocardium was assessed by T1-weigthed Inversion Recovery(T1W-IR), T2-STIR, and T2*Wsequences. All CMR images were matched to histopathology and compared with the area of IMH. The difference between the size of the IMH area detected on T1W-IR images and pathology was −1.6 ± 11.3% (limits of agreement, -24%–21%), for the T2*W images the difference was −0.1 ± 18.3% (limitsof agreement, -36.8%–36.6%), and for T2-STIR the difference was 8.0 ± 15.5% (limits of agreement, -23%–39%). By T1W IR the diagnostic sensitivity of IMH was 90% and specificity 70%, for T2*W imaging the sensitivity was 70% and specificity 50%, and for T2-STIR sensitivity for imaging IMH was 50% and specificity 60%.

Conclusion

T1-weigthednon-contrast enhanced CMR detects IMH with high sensitivity and specificity and may become a diagnostic tool for detection of IMH in patients with myocardial infarction.     

Spatial decision forests for MS lesion segmentation in multi-channel magnetic resonance images

A new algorithm is presented for the automatic segmentation of Multiple Sclerosis (MS) lesions in 3D Magnetic Resonance (MR) images. It builds on a discriminative random decision forest framework to provide a voxel-wise probabilistic classification of the volume. The method uses multi-channel MR intensities (T1, T2, and FLAIR), knowledge on tissue classes and long-range spatial context to discriminate lesions from background. A symmetry feature is introduced accounting for the fact that some MS lesions tend to develop in an asymmetric way. Quantitative evaluation of the proposed methods is carried out on publicly available labeled cases from the MICCAI MS Lesion Segmentation Challenge 2008 dataset. When tested on the same data, the presented method compares favorably to all earlier methods. In an a posteriori analysis, we show how selected features during classification can be ranked according to their discriminative power and reveal the most important ones.     

Segmental high intensity on T1-weighted hepatic MR images

Background We investigated the diagnostic importance of segmental high-intensity (SHI) areas not corresponding to mass lesions on T1-weighted magnetic resonance (MR) images. Methods We conducted a retrospective investigation of hepatic MR images obtained from 634 patients during a 4-year period at our institution. Images were compared with findings reported in the patients’ medical records. There were 16 patients (2.5%) with SHI areas not corresponding to a mass lesion. We compared MR images with plain computed tomographic (CT) scans (n = 16), angiograms (n = 12), and histologic findings (n = 10). Results The segments with intrahepatic bile duct dilatation showed hyperintensity on T1-weighted images. In six of 16 patients, the biliary duct was more dilated in the area of hyperintensity than in areas without hyperintensity. The SHI areas appeared as areas of low attenuation (n = 13), high attenuation (n = 1), or isoattenuation (n = 2) on plain CT scans. Histologically, these areas showed ductular proliferation and deposition of bile pigment within the hepatocytes. Conclusion Segmental areas of increased signal intensity on T1-weighted images were probably due to intrahepatic cholestasis.     

Segmental high intensity on T1-weighted hepatic MR images

BACKGROUND: We investigated the diagnostic importance of segmental high-intensity (SHI) areas not corresponding to mass lesions on T1-weighted magnetic resonance (MR) images. METHODS: We conducted a retrospective investigation of hepatic MR images obtained from 634 patients during a 4-year period at our institution. Images were compared with findings reported in the patients' medical records. There were 16 patients (2.5%) with SHI areas not corresponding to a mass lesion. We compared MR images with plain computed tomographic (CT) scans (n = 16), angiograms (n = 12), and histologic findings (n = 10). RESULTS: The segments with intrahepatic bile duct dilatation showed hyperintensity on T1-weighted images. In six of 16 patients, the biliary duct was more dilated in the area of hyperintensity than in areas without hyperintensity. The SHI areas appeared as areas of low attenuation (n = 13), high attenuation (n = 1), or isoattenuation (n = 2) on plain CT scans. Histologically, these areas showed ductular proliferation and deposition of bile pigment within the hepatocytes. CONCLUSION: Segmental areas of increased signal intensity on T1-weighted images were probably due to intrahepatic cholestasis.     

Automated analysis of four-dimensional magnetic resonance images of the human aorta

The purpose of the study was to demonstrate the accuracy and clinical utility of an automated method of image analysis of 4D (3D + time) magnetic resonance (MR) imaging of the human aorta. Serial MR images of the entire thoracic aorta were acquired on 32 healthy individuals. Graph theory based segmentation was applied to the images and cross sectional area (CSA) was determined for the entire length of thoracic aorta. Mean CSA was compared between the 3 years. CSA values at the level of sinuses of Valsalva and sino-tubular junction were used to calculate average diameters for comparison to Roman-Devereux norms. A robust automated segmentation method was developed that accurately reproduced CSA measurements for the entire length of thoracic aorta in serially acquired scans with a 1% error compared to expert tracing. Calculated aortic root diameters based on CSA correlated with Roman-Devereux norms. Mean CSA for the aortic root agreed well with previously published manually derived values. Automated analysis of 4D MR images of the thoracic aorta provides accurate and reproducible results for CSA in healthy human subjects. The ability to simultaneously analyze the entire length of thoracic aorta throughout the cardiac cycle opens the door to the calculation of novel indices of aortic biophysical properties. These novel indices may lead to earlier detection of patients at risk for adverse events.     

Automated drusen segmentation and quantification in SD-OCT images

Spectral domain optical coherence tomography (SD-OCT) is a useful tool for the visualization of drusen, a retinal abnormality seen in patients with age-related macular degeneration (AMD); however, objective assessment of drusen is thwarted by the lack of a method to robustly quantify these lesions on serial OCT images. Here, we describe an automatic drusen segmentation method for SD-OCT retinal images, which leverages a priori knowledge of normal retinal morphology and anatomical features. The highly reflective and locally connected pixels located below the retinal nerve fiber layer (RNFL) are used to generate a segmentation of the retinal pigment epithelium (RPE) layer. The observed and expected contours of the RPE layer are obtained by interpolating and fitting the shape of the segmented RPE layer, respectively. The areas located between the interpolated and fitted RPE shapes (which have nonzero area when drusen occurs) are marked as drusen. To enhance drusen quantification, we also developed a novel method of retinal projection to generate an en face retinal image based on the RPE extraction, which improves the quality of drusen visualization over the current approach to producing retinal projections from SD-OCT images based on a summed-voxel projection (SVP), and it provides a means of obtaining quantitative features of drusen in the en face projection. Visualization of the segmented drusen is refined through several post-processing steps, drusen detection to eliminate false positive detections on consecutive slices, drusen refinement on a projection view of drusen, and drusen smoothing. Experimental evaluation results demonstrate that our method is effective for drusen segmentation. In a preliminary analysis of the potential clinical utility of our methods, quantitative drusen measurements, such as area and volume, can be correlated with the drusen progression in non-exudative AMD, suggesting that our approach may produce useful quantitative imaging biomarkers to follow this disease and predict patient outcome.     

Automated aorta segmentation in low-dose chest CT images

Purpose

   Abnormalities of aortic surface and aortic diameter can be related to cardiovascular disease and aortic aneurysm. Computer-based aortic segmentation and measurement may aid physicians in related disease diagnosis. This paper presents a fully automated algorithm for aorta segmentation in low-dose non-contrast CT images.

Methods

   The original non-contrast CT scan images as well as their pre-computed anatomy label maps are used to locate the aorta and identify its surface. First a seed point is located inside the aortic lumen. Then, a cylindrical model is progressively fitted to the 3D image space to track the aorta centerline. Finally, the aortic surface is located based on image intensity information. This algorithm has been trained and tested on 359 low-dose non-contrast CT images from VIA-ELCAP and LIDC public image databases. Twenty images were used for training to obtain the optimal set of parameters, while the remaining images were used for testing. The segmentation result has been evaluated both qualitatively and quantitatively. Sixty representative testing images were used to establish a partial ground truth by manual marking on several axial image slices.

Results

   Compared to ground truth marking, the segmentation result had a mean Dice Similarity Coefficient of 0.933 (maximum 0.963 and minimum 0.907). The average boundary distance between manual segmentation and automatic segmentation was 1.39 mm with a maximum of 1.79 mm and a minimum of 0.83 mm.

Conclusion

   Both qualitative and quantitative evaluations have shown that the presented algorithm is able to accurately segment the aorta in low-dose non-contrast CT images.