Automated classification of atherosclerotic plaque from magnetic resonance images using predictive models

The information contained within multicontrast magnetic resonance images (MRI) promises to improve tissue classification accuracy, once appropriately analyzed. Predictive models capture relationships empirically, from known outcomes thereby combining pattern classification with experience. In this study, we examine the applicability of predictive modeling for atherosclerotic plaque component classification of multicontrast ex vivo MR images using stained, histopathological sections as ground truth. Ten multicontrast images from seven human coronary artery specimens were obtained on a 9.4 T imaging system using multicontrast-weighted fast spin-echo (T1-, proton density-, and T2-weighted) imaging with 39-mum isotropic voxel size. Following initial data transformations, predictive modeling focused on automating the identification of specimen's plaque, lipid, and media. The outputs of these three models were used to calculate statistics such as total plaque burden and the ratio of hard plaque (fibrous tissue) to lipid. Both logistic regression and an artificial neural network model (Relevant Input Processor Network-RIPNet) were used for predictive modeling. When compared against segmentation resulting from cluster analysis, the RIPNet models performed between 25 and 30% better in absolute terms. This translates to a 50% higher true positive rate over given levels of false positives. This work indicates that it is feasible to build an automated system of plaque detection using MRI and data mining.     

Comparison of methods for assessing abdominal adipose tissue from magnetic resonance images

Objective: To compare the inter‐rater and intra‐rater reliability and analysis time of two methods for quantifying visceral adipose tissue (VAT) and subcutaneous adipose tissue (SAT) volumes from magnetic resonance (MR) images. Research Methods and Procedures: Ten subjects (BMI, 27.0 ± 2.1 kg/m²; 56 years of age ± 4 years) underwent MR imaging of the abdomen. Ten transverse T1‐weighted images were selected from each scan and analyzed using two software packages that differ in principle. The first method, ANALYZE version 5.0, represents the manual threshold method, and the second, HIPPO version 1.3, is based on the fuzzy clustering approach. Inter‐rater reliability for each method was assessed by comparing the intra‐class correlation coefficients (ICCs) for VAT and SAT results from two evaluators, and intra‐rater reliability for each method was assessed by comparing ICCs for VAT and SAT analyses performed 1 week apart by the same evaluator. The total time for analysis also was compared between methods. Results: The inter‐rater reliability for VAT was greater with HIPPO than with ANALYZE (ICC = 0.996 vs. 0.828), whereas inter‐rater reliability for SAT did not differ between methods (ICC = 0.975 and 0.987). The intra‐rater reliability was equally high with HIPPO and ANALYZE for both VAT (ICC = 0.998 vs. 0.992) and SAT (ICC = 0.996 vs. 0.992). HIPPO required less than one‐half as much analysis time as ANALYZE (15.9 ± 4.4 vs. 36.5 ± 8.2 minutes, p < 0.0001). Discussion: HIPPO software appears advantageous for the quantification of VAT from multislice MR images because inter‐rater results are more reliable, and it is more time‐efficient than less automated methods.     

Automated analysis of craniofacial morphology using magnetic resonance images

Quantitative analysis of craniofacial morphology is of interest to scholars working in a wide variety of disciplines, such as anthropology, developmental biology, and medicine. T1-weighted (anatomical) magnetic resonance images (MRI) provide excellent contrast between soft tissues. Given its three-dimensional nature, MRI represents an ideal imaging modality for the analysis of craniofacial structure in living individuals. Here we describe how T1-weighted MR images, acquired to examine brain anatomy, can also be used to analyze facial features. Using a sample of typically developing adolescents from the Saguenay Youth Study (N = 597; 292 male, 305 female, ages: 12 to 18 years), we quantified inter-individual variations in craniofacial structure in two ways. First, we adapted existing nonlinear registration-based morphological techniques to generate iteratively a group-wise population average of craniofacial features. The nonlinear transformations were used to map the craniofacial structure of each individual to the population average. Using voxel-wise measures of expansion and contraction, we then examined the effects of sex and age on inter-individual variations in facial features. Second, we employed a landmark-based approach to quantify variations in face surfaces. This approach involves: (a) placing 56 landmarks (forehead, nose, lips, jaw-line, cheekbones, and eyes) on a surface representation of the MRI-based group average; (b) warping the landmarks to the individual faces using the inverse nonlinear transformation estimated for each person; and (3) using a principal components analysis (PCA) of the warped landmarks to identify facial features (i.e. clusters of landmarks) that vary in our sample in a correlated fashion. As with the voxel-wise analysis of the deformation fields, we examined the effects of sex and age on the PCA-derived spatial relationships between facial features. Both methods demonstrated significant sexual dimorphism in craniofacial structure in areas such as the chin, mandible, lips, and nose.     

Estimating whole body intermuscular adipose tissue from single cross-sectional magnetic resonance images.

Intermuscular adipose tissue (IMAT), a novel fat depot linked with metabolic abnormalities, has been measured by whole body MRI. The cross-sectional slice location with the strongest relation to total body IMAT volume has not been established. The aim was to determine the predictive value of each slice location and which slice locations provide the best estimates of whole body IMAT. MRI quantified total adipose tissue of which IMAT, defined as adipose tissue visible within the boundary of the muscle fascia, is a subcomponent. Single-slice IMAT areas were calculated for the calf, thigh, buttock, waist, shoulders, upper arm, and forearm locations in a sample of healthy adult women, African-American [n = 39; body mass index (BMI) 28.5 +/- 5.4 kg/m2; 41.8 +/- 14.8 yr], Asian (n = 21; BMI 21.6 +/- 3.2 kg/m2; 40.9 +/- 16.3 yr), and Caucasian (n = 43; BMI 25.6 +/- 5.3 kg/m2; 43.2 +/- 15.3 yr), and Caucasian men (n = 39; BMI 27.1 +/- 3.8 kg/m2; 45.2 +/- 14.6 yr) and used to estimate total IMAT groups using multiple-regression equations. Midthigh was the best, or near best, single predictor in all groups with adjusted R2 ranging from 0.49 to 0.84. Adding a second and third slice further increased R2 and reduced the error of the estimate. Menopausal status and degree of obesity did not affect the location of the best single slice. The contributions of other slice locations varied by sex and race, but additional slices improved predictions. For group studies, it may be more cost-effective to estimate IMAT based on one or more slices than to acquire and segment for each subject the numerous images necessary to quantify whole body IMAT.     

Data distributions in magnetic resonance images: A review

Many image processing methods applied to magnetic resonance (MR) images directly or indirectly rely on prior knowledge of the statistical data distribution that characterizes the MR data. Also, data distributions are key in many parameter estimation problems and strongly relate to the accuracy and precision with which parameters can be estimated. This review paper provides an overview of the various distributions that occur when dealing with MR data, considering both single-coil and multiple-coil acquisition systems. The paper also summarizes how knowledge of the MR data distributions can be used to construct optimal parameter estimators and answers the question as to what precision may be achieved ultimately from a particular MR image.     

Improvement of functional magnetic resonance images by pretreatment of data

Functional magnetic resonance images of the brains of subjects performing the finger-tapping paradigm were made using a conventional technique. Two threshold values for the pixels were obtained by analysing pixel by pixel the distributions of the means and variances of each subject's images for 20 consecutive scans, both while performing the task and while at rest. Considerable signal improvement in the final images was achieved by removing from our data all pixels beyond these threshold values (mean 16 and variance 7).     

New head models extracted from thermal infrared (IR) images for dosimetry computations

In electromagnetic dosimetry, anatomical human models are commonly obtained by segmentation of magnetic resonance imaging or computed tomography scans. In this paper, a human head model extracted from thermal infrared images is examined in terms of its applicability to specific absorption rate (SAR) calculations. Since thermal scans are two-dimensional (2D) representation of surface temperature, this allows researchers to overcome the extensive computational demand associated with 3D simulation. The numerical calculations are performed using the finite-difference time-domain method with mesh sizes of 2 mm at 900 MHz plane wave irradiation. The power density of the incident plane wave is assumed to be 10 W/m². Computations were compared with a realistic anatomical head model. The results show that although there were marked differences in the local SAR distribution in the various tissues in the two models, the 1 g peak SAR values are approximately similar in the two models.     

New head models extracted from thermal infrared (IR) images for dosimetry computations

In electromagnetic dosimetry, anatomical human models are commonly obtained by segmentation of magnetic resonance imaging or computed tomography scans. In this paper, a human head model extracted from thermal infrared images is examined in terms of its applicability to specific absorption rate (SAR) calculations. Since thermal scans are two-dimensional (2D) representation of surface temperature, this allows researchers to overcome the extensive computational demand associated with 3D simulation. The numerical calculations are performed using the finite-difference time-domain method with mesh sizes of 2 mm at 900 MHz plane wave irradiation. The power density of the incident plane wave is assumed to be 10 W/m(2). Computations were compared with a realistic anatomical head model. The results show that although there were marked differences in the local SAR distribution in the various tissues in the two models, the 1 g peak SAR values are approximately similar in the two models.     

Hypothalamic hamartoma: comparison of clinical presentation and magnetic resonance images

BACKGROUND/AIMS: Hypothalamic hamartoma (HH) is one of the most frequent causes of organic central precocious puberty (CPP). We compared the clinical presentation and the magnetic resonance images (MRI) of 19 patients with HH aged 5.7 +/- 4.1 (SD) years at the first endocrine evaluation. They had isolated CPP (group 1, n = 9), CPP plus gelastic seizures (group 2, n = 5), isolated seizures (group 3, n = 4), and 1 patient was asymptomatic. METHODS/RESULTS: All patients without neurological symptoms (group 1 and the asymptomatic patient) had pedunculated lesion (diameter 6.4 +/- 3.6 (3-15) mm), suspended from the floor of the third ventricle. All patients with neurological symptoms (groups 2 and 3) had sessile lesion (diameter 18.3 +/- 9.6 (10-38) mm, p = 0.0005 compared to the others), located in the interpeduncular cistern with extension to the hypothalamus. Seven patients were overweight. The growth hormone peak, free thyroxine, cortisol and prolactin concentrations, and the concomitant plasma and urinary osmolalities were normal in all the cases evaluated. The mean predicted or adult heights of 10 patients treated 5.2 +/- 3.3 years for CPP with gonadotropin hormone releasing hormone (GnRH) analog were -0.3 +/- 1.7 SD, similar to their target height -0.1 +/- 0.9 SD. CONCLUSION: The clinical presentation of HH depends on its anatomy: small and pedunculated HH are associated with CPP, while large and sessile HH are associated with seizures. The hypothalamic-pituitary function in these cases is normal, which suggests that the absence of CPP is not due to gonadotropin deficiency. GnRH analog treatment preserves the growth potential in those with CPP.     

Nuclear magnetic resonance characterization of peptide models of collagen-folding diseases

Misfolding of the triple helix has been shown to play a critical role in collagen diseases. The substitution of a single Gly by another amino acid breaks the characteristic repeating (Gly-X-Y)n sequence pattern and results in connective tissue disease such as osteogenesis imperfecta. Nuclear magnetic resonance (NMR) studies of normal and mutated collagen triple-helical peptides offer an opportunity to characterize folding and conformational alterations at the substitution site, as well as at positions upstream and downstream of a Gly mutation. The NMR studies suggest that the local sequences surrounding the substitution site, and the renucleation sequences N-terminal to and adjacent to the substitution site, may be critical in defining the clinical phenotype of osteogenesis imperfecta. These studies may pave the way to understanding the mechanism by which a single Gly substitution in collagen can lead to pathological conditions.     

Quantification of meniscal volume by segmentation of 3T magnetic resonance images

Meniscal injuries place the knee at risk for early osteoarthritis (OA) because they disrupt their load-bearing capabilities. Partial resection is routinely performed to alleviate symptomatic meniscal tears. While the removal of meniscal tissue may not be the only factor associated with partial meniscectomy outcome, the amount removed certainly contributes to functional loss. It is unknown, however, whether there is a critical amount of meniscal tissue that can be removed without diminishing the structure's chondroprotective role. In order to examine the existence of such a threshold, it is necessary to accurately quantify meniscal volume both before and after partial meniscectomy to determine the amount of meniscal tissue removed. Therefore, our goal was to develop and validate an MR-based method for assessing meniscal volume. The specific aims were: (1) to evaluate the feasibility of the MR-based segmentation method; (2) to determine the method's reliability for repeated measurements; and (3) to validate its accuracy in situ. MR images were obtained on a 3T magnet, and each scan was segmented using a biplanar approach. The MR-based volumes for each specimen were compared to those measured by water displacement. The results indicate that the biplanar approach of measuring meniscal volumes is accurate and reliable. The calculated volumes of the menisci were within 5% of the true values, the coefficients of variation were 4%, and the intraclass correlation coefficients were greater than 0.96. These data demonstrate that this method could be used to measure the amount of meniscal tissue excised during partial meniscectomy to within 125.7 mm(3).     

Pixel T2 distribution in functional magnetic resonance images of muscle.

Increases in skeletal muscle (1)H-NMR transverse relaxation time (T2) observed by magnetic resonance imaging have been used to map whole muscle activity during exercise. Some studies further suggest that intramuscular variations in T2 after exercise can be used to map activity on a pixel-by-pixel basis by defining an active T2 threshold and counting pixels that exceed the threshold as "active muscle." This implies that motor units are nonrandomly distributed across the muscle and, therefore, that the distribution of pixel T2 values ought to be substantially broader after moderate exercise than at rest or after more intense exercise, since moderate-intensity exercise should recruit some motor units, and hence some pixels, but not others. This study examined the distribution of pixel T2 values in three muscles (quadriceps, anterior tibialis, and biceps/brachialis) of healthy subjects (5 men and 2 women, 18-46 yr old) at rest, after exercise to fatigue (50% 1 repetition maximum at 20/min to failure = Max), and at 1/2Max (25% 1 repetition maximum, same number of repetitions as Max). Although for each muscle there was a linear relationship between exercise intensity and mean pixel T2, there was no significant difference in the variance of pixel T2 between 1/2Max and Max exercise. There was a modest (10-43%) increase in variance of pixel T2 after both exercises compared with rest, but this was consistent with a Monte Carlo simulation of muscle activity that assumed a random distribution of motor unit territories across the muscle and a random distribution of muscle cells within each motor unit's territory. In addition, 40% of the pixel-to-pixel muscle T2 variations were shown to be due to imaging noise. The results indicate that magnetic resonance imaging T2 cannot reliably map active muscle on a pixel-by-pixel basis in normal subjects.     

Construction of hydrodynamic bead models from high-resolution X-ray crystallographic or nuclear magnetic resonance data.

Computer software such as HYDRO, based upon a comprehensive body of theoretical work, permits the hydrodynamic modeling of macromolecules in solution, which are represented to the computer interface as an assembly of spheres. The uniqueness of any satisfactory resultant model is optimized by incorporating into the modeling procedure the maximal possible number of criteria to which the bead model must conform. An algorithm (AtoB, for atoms to beads) that permits the direct construction of bead models from high resolution x-ray crystallographic or nuclear magnetic resonance data has now been formulated and tested. Models so generated then act as informed starting estimates for the subsequent iterative modeling procedure, thereby hastening the convergence to reasonable representations of solution conformation. Successful application of this algorithm to several proteins shows that predictions of hydrodynamic parameters, including those concerning solvation, can be confirmed.     

Velocity profiles in stenosed tube models using magnetic resonance imaging

A time-of-flight MRI velocity measurement technique is evaluated against corresponding LDV measurements in a constriction tube model over a range of physiologic flow conditions. Results from this study show that MR displacement images can: 1) be obtained within both laminar and turbulent jets (maximum stenotic Re approximately equal to 4,200), 2) measure mean jet velocities up to 172 cm/s, and, 3) detect low forward and reverse stenosis (0 less than or equal to L/D less than or equal to 2). Regions between the jet termination point and re-establishment of laminar flow (Re greater than or equal to 1500, greater than or equal to 1000, and greater than or equal to 110 downstream of 40, 60, and 80 percent stenosis, respectively) cannot presently be detected by this technique.     

Changes in magnetic resonance images in human skeletal muscle after eccentric exercise

To investigate the time-course of changes in transverse relaxation time (T₂) and cross-sectional area (CSA) of the quadriceps muscle after a single session of eccentric exercise, magnetic resonance imaging was performed on six healthy male volunteers before and at 0, 7, 15, 20, 30 and 60 min and 12, 24, 36, 48, 72 and 168 h after exercise. Although there was almost no muscle soreness immediately after exercise, it started to increase 1 day after, peaking 1–2 days after the exercise (P<0.01). Immediately after exercise, T₂ increased significantly in the rectus femoris, vastus lateralis and intermedius muscles (P<0.05) and decreased quickly continuing until 60 min after exercise. At and after the 12th h, a significant increase was perceived again in the T₂ values of the vastus lateralis and intermedius muscles (P<0.01) [maximum 9.3 (SEM 2.8)% and 10.9 (SEM 2.2)%, respectively]. The maximal values were exhibited at 24–36 h after exercise. In contrast, the rectus femoris muscle showed no delayed-stage increase. Also, in CSA, an increase after 12 h was observed in addition to the one immediately after exercise in the vastus lateralis, intermedius and medialis and quadriceps muscles as a whole (P < 0.01), reaching the maximal values at 12–24 h after exercise. The plasma creative kinase activity remained unchanged up to 24 h after and then increased significantly 48 h after exercise (P < 0.05). Beginning 12 h after exercise, the subjects whose T₂ and CSA increased less than the others displayed a faster decrease in muscle soreness. These results suggested that T₂ and CSA displayed bimodal responses after eccentric exercise and the time-courses of changes in them were similar to those in muscle soreness.