Quantification of multiple sclerosis lesion volumes in 1.5 and 0.5 T anisotropically filtered and unfiltered MR exams

Deep neural network to locate and segment brain tumors outperformed the expert technicians who created the training data

Purpose: Deep learning (DL) algorithms have shown promising results for brain tumor segmentation in MRI. However, validation is required prior to routine clinical use. We report the first randomized and blinded comparison of DL and trained technician segmentations.

Approach: We compiled a multi-institutional database of 741 pretreatment MRI exams. Each contained a postcontrast T1-weighted exam, a T2-weighted fluid-attenuated inversion recovery exam, and at least one technician-derived tumor segmentation. The database included 729 unique patients (470 males and 259 females). Of these exams, 641 were used for training the DL system, and 100 were reserved for testing. We developed a platform to enable qualitative, blinded, controlled assessment of lesion segmentations made by technicians and the DL method. On this platform, 20 neuroradiologists performed 400 side-by-side comparisons of segmentations on 100 test cases. They scored each segmentation between 0 (poor) and 10 (perfect). Agreement between segmentations from technicians and the DL method was also evaluated quantitatively using the Dice coefficient, which produces values between 0 (no overlap) and 1 (perfect overlap).

Results: The neuroradiologists gave technician and DL segmentations mean scores of 6.97 and 7.31, respectively (p<0.00007). The DL method achieved a mean Dice coefficient of 0.87 on the test cases.

Conclusions: This was the first objective comparison of automated and human segmentation using a blinded controlled assessment study. Our DL system learned to outperform its “human teachers” and produced output that was better, on average, than its training data.

Clinical evaluation of stereotactic target localization using 3-Tesla MRI for radiosurgery planning

PURPOSE

Increasing the magnetic resonance imaging (MRI) field strength can improve image resolution and quality, but concerns remain regarding the influence on geometric fidelity. The objectives of the present study were to spatially investigate the effect of 3-Tesla (3T) MRI on clinical target localization for stereotactic radiosurgery.

METHODS AND MATERIALS

A total of 39 patients were enrolled in a research ethics board-approved prospective clinical trial. Imaging (1.5T and 3T MRI and computed tomography) was performed after stereotactic frame placement. Stereotactic target localization at 1.5T vs. 3T was retrospectively analyzed in a representative cohort of patients with tumor (n = 4) and functional (n = 5) radiosurgical targets. The spatial congruency of the tumor gross target volumes was determined by the mean discrepancy between the average gross target volume surfaces at 1.5T and 3T. Reproducibility was assessed by the displacement from an averaged surface and volume congruency. Spatial congruency and the reproducibility of functional radiosurgical targets was determined by comparing the mean and standard deviation of the isocenter coordinates.

RESULTS

Overall, the mean absolute discrepancy across all patients was 0.67 mm (95% confidence interval, 0.51-0.83), significantly <1 mm (p < .010). No differences were found in the overall interuser target volume congruence (mean, 84% for 1.5T vs. 84% for 3T, p > .4), and the gross target volume surface mean displacements were similar within and between users. The overall average isocenter coordinate discrepancy for the functional targets at 1.5T and 3T was 0.33 mm (95% confidence interval, 0.20-0.48), with no patient-specific differences between the mean values (p >.2) or standard deviations (p >.1).

CONCLUSION

Our results have provided clinically relevant evidence supporting the spatial validity of 3T MRI for use in stereotactic radiosurgery under the imaging conditions used.

Ovarian volume measurements in mice with high-resolution ultrasonography

OBJECTIVE

The aim of our study was to evaluate the intraobserver and interobserver variability of ovarian volume measurements in mice with high-resolution 2-dimensional ultrasonography (2DUS) and 3-dimensional ultrasonography (3DUS).

METHODS

Ovaries of 10 nude mice were visualized with a small-animal ultrasound scanner and a 40-MHz probe. For each ovary, volume was measured 3 times by 2 independent readers using both 2DUS and 3DUS methods. The 2DUS method used a biplane ellipsoid model. The 3DUS method estimated the volume by integrating 10 to 12 parallel image planes of the ovary after semiautomated outlining of the boundaries. For each type of measurement, intraobserver and interobserver standard error of measurement (SEM) values and minimal detectable volume changes were calculated by analysis of variance.

RESULTS

Two-dimensional ultrasonography showed much poorer reproducibility, with higher absolute intraobserver and interobserver SEM values (0.50 and 0.61 mm3, respectively) than 3DUS (0.20 and 0.35 mm3; P < .01). Relative intraobserver and interobserver SEM values were also much higher for 2DUS (12.20% and 14.88%) than for 3DUS (5.12% and 8.97%; P < .01). The minimal volume changes that could be detected with a 95% confidence level in successive measurements by the same (or different) observers were 33.90% (41.22%) for 2DUS and 14.10% (24.87%) for 3DUS.

CONCLUSIONS

High-resolution 3DUS can provide a reliable tool for noninvasive, longitudinal ovarian volume measurements in mice.

Intraobserver and interobserver variability in GTV delineation on FDG-PET-CT images of head and neck cancers

PURPOSE

To determine if the addition of fluorodeoxyglucose positron emission tomography (FDG-PET) data changes primary site gross tumor volumes (GTVs) in head and neck cancers.

METHODS AND MATERIALS

Computed tomography (CT), contrast-enhanced CT, and FDG-PET-CT scans were obtained in 10 patients with head and neck cancers. Eight experienced observers (6 head and neck oncologists and 2 neuro-radiologists) with access to clinical and radiologic reports outlined primary site GTVs on each modality. Three cases were recontoured twice to assess intraobserver variability. The magnitudes of the GTVs were compared. Intra- and interobserver variability was assessed by a two-way repeated measures analysis of variance. Inter- and intraobserver reliability were calculated.

RESULTS

There were no significant differences in the GTVs across the image modalities when compared as ensemble averages; the Wilcoxon matched-pairs signed-rank test showed that CT volumes were larger than PET-CT. Observers demonstrated the greatest consistency and were most interchangeable on contrast-enhanced CT; they performed less reliably on PET-CT.

CONCLUSIONS

The addition of PET-CT to primary site GTV delineation of head and neck cancers does not change the volume of the GTV defined by this group of expert observers in this patient sample. An FDG-PET may demonstrate differences in neck node delineation and in other disease sites.

Multiple sclerosis medical image analysis and information management

Magnetic resonance imaging (MRI) has become a central tool for patient management, as well as research, in multiple sclerosis (MS). Measurements of disease burden and activity derived from MRI through quantitative image analysis techniques are increasingly being used. There are many complexities and challenges in building computerized processing pipelines to ensure efficiency, reproducibility, and quality control for MRI scans from MS patients. Such paradigms require advanced image processing and analysis technologies, as well as integrated database management systems to ensure the most utility for clinical and research purposes. This article reviews pipelines available for quantitative clinical MRI research in MS, including image segmentation, registration, time-series analysis, performance validation, visualization techniques, and advanced medical imaging software packages. To address the complex demands of the sequential processes, the authors developed a workflow management system that uses a centralized database and distributed computing system for image processing and analysis. The implementation of their system includes a web-form-based Oracle database application for information management and event dispatching, and multiple modules for image processing and analysis. The seamless integration of processing pipelines with the database makes it more efficient for users to navigate complex, multistep analysis protocols, reduces the user's learning curve, reduces the time needed for combining and activating different computing modules, and allows for close monitoring for quality-control purposes. The authors' system can be extended to general applications in clinical trials and to routine processing for image-based clinical research.

Robust myelin water quantification: averaging vs. spatial filtering

The myelin water fraction is calculated, voxel-by-voxel, by fitting decay curves from a multi-echo data acquisition. Curve-fitting algorithms require a high signal-to-noise ratio to separate T(2) components in the T(2) distribution. This work compared the effect of averaging, during acquisition, to data postprocessed with a noise reduction filter. Forty regions, from five volunteers, were analyzed. A consistent decrease in the myelin water fraction variability with no bias in the mean was found for all 40 regions. Images of the myelin water fraction of white matter were more contiguous and had fewer "holes" than images of myelin water fractions from unfiltered echoes. Spatial filtering was effective for decreasing the variability in myelin water fraction calculated from 4-average multi-echo data.

Automatic "pipeline" analysis of 3-D MRI data for clinical trials: application to multiple sclerosis

The quantitative analysis of magnetic resonance imaging (MRI) data has become increasingly important in both research and clinical studies aiming at human brain development, function, and pathology. Inevitably, the role of quantitative image analysis in the evaluation of drug therapy will increase, driven in part by requirements imposed by regulatory agencies. However, the prohibitive length of time involved and the significant intraand inter-rater variability of the measurements obtained from manual analysis of large MRI databases represent major obstacles to the wider application of quantitative MRI analysis. We have developed a fully automatic "pipeline" image analysis framework and have successfully applied it to a number of large-scale, multicenter studies (more than 1,000 MRI scans). This pipeline system is based on robust image processing algorithms, executed in a parallel, distributed fashion. This paper describes the application of this system to the automatic quantification of multiple sclerosis lesion load in MRI, in the context of a phase III clinical trial. The pipeline results were evaluated through an extensive validation study, revealing that the obtained lesion measurements are statistically indistinguishable from those obtained by trained human observers. Given that intra- and inter-rater measurement variability is eliminated by automatic analysis, this system enhances the ability to detect small treatment effects not readily detectable through conventional analysis techniques. While useful for clinical trial analysis in multiple sclerosis, this system holds widespread potential for applications in other neurological disorders, as well as for the study of neurobiology in general.

3-D surface reconstruction of multiple sclerosis lesions using spherical harmonics

A new approach to approximate the 3-D shape of multiple sclerosis (MS) lesions and to calculate their volumes is presented. The suggested method utilizes sets of MS lesion contours taken from segmented MR images and approximates their 3-D surfaces by spherical harmonics. This method was applied to obtain 3-D reconstructions of in vivo and simulated MS lesions and to calculate their volumes. The results show good geometrical approximations of the original MS lesions' 3-D shapes and good consistency in volume estimation independent of the size of the lesions. The average volume estimation error was smaller than the commonly used technique of slice stacking (15.5 +/- 13.4% and 13.1 +/- 10.1% vs. 25.0 +/- 17.0%). The method presented here offers a tool for analyzing the geometrical characteristics of MS lesions in 3-D as well as their volumes. The geometrical information may potentially serve as an additional clinical index for monitoring the disease.

Effect of black blood MR image quality on vessel wall segmentation

Black blood MRI has become a popular technique for measuring arterial wall area as an indicator of plaque size. Computer-assisted techniques for segmenting vessel boundaries have been developed to increase measurement precision. In this study, the carotid arteries of four normal subjects were imaged at seven different fields of view (FOVs), keeping all other imaging parameters fixed, to determine whether spatial resolution could be increased at the expense of image quality without sacrificing precision. Wall areas were measured via computer-assisted segmentation of the vessel boundaries performed repeatedly by two operators. Analysis of variance (ANOVA) demonstrated that the variability of wall area measurements was below 1.5 mm(2) for in-plane spatial resolutions between 0.22 mm and 0.37 mm. An inverse relationship between operator variability and the signal difference-to-noise ratio (SDNR) demonstrated that semi-automatic segmentation of the wall boundaries was robust for SDNR >3, defining a criterion above which subjective image quality can be degraded without an appreciable loss of information content. Our study also suggested that spatial resolutions higher than 0.3 mm may be required to quantify normal wall areas to within 10% accuracy, but that the reduced SNR associated with the higher resolution may be tolerated by semi-automated wall segmentation without an appreciable loss of precision.

A semi-automatic technique for measurement of arterial wall from black blood MRI

Black blood magnetic resonance imaging (MRI) has become a popular technique for imaging the artery wall in vivo. Its noninvasiveness and high resolution make it ideal for studying the progression of early atherosclerosis in normal volunteers or asymptomatic patients with mild disease. However, the operator variability inherent in the manual measurement of vessel wall area from MR images hinders the reliable detection of relatively small changes in the artery wall over time. In this paper we present a semi-automatic method for segmenting the inner and outer boundary of the artery wall, and evaluate its operator variability using analysis of variance (ANOVA). In our approach, a discrete dynamic contour is approximately initialized by an operator, deformed to the inner boundary, dilated, and then deformed to the outer boundary. A group of four operators performed repeated measurements on 12 images from normal human subjects using both our semiautomatic technique and a manual approach. Results from the ANOVA indicate that the inter-operator standard error of measurement (SEM) of total wall area decreased from 3.254 mm2 (manual) to 1.293 mm2 (semi-automatic), and the intra-operator SEM decreased from 3.005 mm2 to 0.958 mm2. Operator reliability coefficients increased from less than 69% to more than 91% (inter-operator) and 95% (intra-operator). The minimum detectable change in wall area improved from more than 8.32 mm2 (intra-operator, manual) to less than 3.59 mm2 (inter-operator, semi-automatic), suggesting that it is better to have multiple operators measure wall area with our semi-automatic technique than to have a single operator make repeated measurements manually. Similar improvements in wall thickness and lumen radius measurements were also recorded. Since the semi-automatic technique has effectively ruled out the effect of the operator on these measurements, it may be possible to use such techniques to expand prospective studies of atherogenesis to multiple centers so as to increase access to real patient data without sacrificing reliability.

Quantitative follow-up of patients with multiple sclerosis using MRI: reproducibility

The reproducibility of an automated method for estimating the volume of white matter abnormalities on brain magnetic resonance (MR) images of multiple sclerosis (MS) patients was evaluated. Twenty MS patients underwent MR imaging twice within 30 minutes. Measurement variability is introduced mainly by MRI acquisition and image registration procedures, which demonstrate significantly worse reproducibility than the image segmentation. The correction of partial volume artifacts is essential for sensitive measurements of overall lesion burden. The average lesion volume difference (bias) between two MR exams of the same MS patient (N = 20) was 0.05 cm3, with a 95% confidence interval between -0.17 and +0.28 cm3, suggesting that the proposed measurement system is suitable for clinical follow-up trials, even in relatively small patient cohorts. The limits of agreement for lesion volume were between -1.3 and +1.5 cm3, implying that in individual patients changes in lesion load need to be at least this large to be detected reliably. This automated method for estimating lesion burden is a reliable tool for the evaluation of MS progression and exacerbation in patient cohorts and potentially also in individual patients.

Selective measurement of white matter and gray matter diffusion trace values in normal human brain

The trace of the diffusion tensor (or simply the trace) is diagnostically valuable for detecting acute ischemic lesions. A number of studies indicate that the trace of human gray matter (GM) and white matter (WM) are quite similar. This is somewhat surprising considering the different cellular environments of GM and WM. It is possible that partial volume averaging (PVA) effects between GM and WM, inherent in many of the ultrafast imaging sequences used for diffusion measurements, are responsible for this observation. In order to minimize PVA effects, the trace values of GM and WM have been selectively measured by implementing double inversion recovery (DIR) echo planar imaging (EPI) pulse sequences. Results on six normal volunteers indicate that the trace values of WM and GM are not statistically different.

Intra- and inter-observer variability and reliability of prostate volume measurement via two-dimensional and three-dimensional ultrasound imaging

We describe the results of a study to evaluate the intra- and inter-observer variability and reliability of prostate volume measurements made from transrectal ultrasound (TRUS) images, using either the (optimal) height-width-length (HWL) method (V = pi/6 HWL) with two-dimensional (2D) TRUS images (obtained as cross-sections of three-dimensional [3D] TRUS images) or manual planimetry of 3D TRUS images (the 3D US method). In this study, eight observers measured 15 prostate images, twice via each method, and an analysis of variance (ANOVA) was performed. This analysis shows that, with the 3D US method, intra-observer prostate volume estimates have 5.1% variability and 99% reliability, and inter-observer estimates have 11.4% variability and 96% reliability. With the HWL method, intra-observer estimates have 15.5% variability and 93% reliability, and inter-observer estimates have 21.9% variability and 87% reliability. Thus, in vivo prostate volume estimates from manual planimetry of 3D TRUS images have much lower variability and higher reliability than HWL estimates from 2D TRUS images.

A dual approach for minimizing false lesion classifications on magnetic resonance images

Segmentation methods based on dual-echo MR images are generally prone to significant false lesion classifications. We have minimized these false classifications by (1) improving the lesion-to-tissue contrast on MR images by developing a fast spin-echo sequence that incorporates both cerebrospinal fluid signal attenuation and magnetization transfer contrast and (2) including information from MR flow images. Studies on patients with multiple sclerosis indicate that this dual approach to tissue segmentation reduces the volume of false lesion classifications by an average of 87%.

Cerebrospinal fluid-suppressed high-resolution diffusion imaging of human brain

A cerebrospinal fluid (CSF)-suppressed flow-attenuated inversion recovery (FLAIR) double-shot diffusion echo-planar imaging (EPI) sequence was developed and used, along with a non-CSF-suppressed version of the sequence, to determine the extent of the contribution of CSF partial-volume averaging to the apparent diffusion coefficients (ADCs) of normal human brain in vivo. Regional analysis indicates that cortical gray matter and parenchymal tissues bordering the ventricles are most affected by CSF contamination, leading to elevated ADC values. Only slight differences in gray- and white-matter average ADCs were detected after CSF suppression. The human brain average ADCs calculated from high-resolution CSF-suppressed diffusion-weighted images in these studies are similar to those reported in animals. FLAIR diffusion sequences remove CSF as a source of error in ADC determination and ischemic lesion discrimination in diffusion-weighted images (DWI) and ADC maps.