Optimization of MRI protocols and pulse sequence parameters for eigenimage filtering

AI-driven and automated MRI sequence optimization in scanner-independent MRI sequences formulated by a domain-specific language

Introduction

The complexity of Magnetic Resonance Imaging (MRI) sequences requires expert knowledge about the underlying contrast mechanisms to select from the wide range of available applications and protocols. Automation of this process using machine learning (ML) can support the radiologists and MR technicians by complementing their experience and finding the optimal MRI sequence and protocol for certain applications.

Methods

We define domain-specific languages (DSL) both for describing MRI sequences and for formulating clinical demands for sequence optimization. By using various abstraction levels, we allow different key users exact definitions of MRI sequences and make them more accessible to ML. We use a vendor-independent MRI framework (gammaSTAR) to build sequences that are formulated by the DSL and export them using the generic file format introduced by the Pulseq framework, making it possible to simulate phantom data using the open-source MR simulation framework JEMRIS to build a training database that relates input MRI sequences to output sets of metrics. Utilizing ML techniques, we learn this correspondence to allow efficient optimization of MRI sequences meeting the clinical demands formulated as a starting point.

Results

ML methods are capable of capturing the relation of input and simulated output parameters. Evolutionary algorithms show promising results in finding optimal MRI sequences with regards to the training data. Simulated and acquired MRI data show high correspondence to the initial set of requirements.

Discussion

This work has the potential to offer optimal solutions for different clinical scenarios, potentially reducing exam times by preventing suboptimal MRI protocol settings. Future work needs to cover additional DSL layers of higher flexibility as well as an optimization of the underlying MRI simulation process together with an extension of the optimization method.

Autonomous magnetic resonance imaging

Access to Magnetic Resonance Imaging (MRI) across developing countries ranges from being prohibitive to scarcely available. For example, eleven countries in Africa have no scanners. One critical limitation is the absence of skilled manpower required for MRI usage. Some of these challenges can be mitigated using autonomous MRI (AMRI) operation. In this work, we demonstrate AMRI to simplify MRI workflow by separating the required intelligence and user interaction from the acquisition hardware. AMRI consists of three components: user node, cloud and scanner. The user node voice interacts with the user and presents the image reconstructions at the end of the AMRI exam. The cloud generates pulse sequences and performs image reconstructions while the scanner acquires the raw data. An AMRI exam is a custom brain screen protocol comprising of one T₁-, T₂- and T₂*-weighted exams. A neural network is trained to incorporate Intelligent Slice Planning (ISP) at the start of the AMRI exam. A Look Up Table was designed to perform intelligent protocolling by optimizing for contrast value while satisfying signal to noise ratio and acquisition time constraints. Data were acquired from four healthy volunteers for three experiments with different acquisition time constraints to demonstrate standard and self-administered AMRI. The source code is available online. AMRI achieved an average SNR of 22.86 ± 0.89 dB across all experiments with similar contrast. Experiment #3 (33.66% shorter table time than experiment #1) yielded a SNR of 21.84 ± 6.36 dB compared to 23.48 ± 7.95 dB for experiment #1. AMRI can potentially enable multiple scenarios to facilitate rapid prototyping and research and streamline radiological workflow. We believe we have demonstrated the first Autonomous MRI of the brain.

Multiparametric deep learning tissue signatures for a radiological biomarker of breast cancer: Preliminary results

PURPOSE

Deep learning is emerging in radiology due to the increased computational capabilities available to reading rooms. These computational developments have the ability to mimic the radiologist and may allow for more accurate tissue characterization of normal and pathological lesion tissue to assist radiologists in defining different diseases. We introduce a novel tissue signature model based on tissue characteristics in breast tissue from multiparametric magnetic resonance imaging (mpMRI). The breast tissue signatures are used as inputs in a stacked sparse autoencoder (SSAE) multiparametric deep learning (MPDL) network for segmentation of breast mpMRI.

METHODS

We constructed the MPDL network from SSAE with 5 layers with 10 nodes at each layer. A total cohort of 195 breast cancer subjects were used for training and testing of the MPDL network. The cohort consisted of a training dataset of 145 subjects and an independent validation set of 50 subjects. After segmentation, we used a combined SAE-support vector machine (SAE-SVM) learning method for classification. Dice similarity (DS) metrics were calculated between the segmented MPDL and dynamic contrast enhancement (DCE) MRI-defined lesions. Sensitivity, specificity, and area under the curve (AUC) metrics were used to classify benign from malignant lesions.

RESULTS

The MPDL segmentation resulted in a high DS of 0.87 ± 0.05 for malignant lesions and 0.84 ± 0.07 for benign lesions. The MPDL had excellent sensitivity and specificity of 86% and 86% with positive predictive and negative predictive values of 92% and 73%, respectively, and an AUC of 0.90.

CONCLUSIONS

Using a new tissue signature model as inputs into the MPDL algorithm, we have successfully validated MPDL in a large cohort of subjects and achieved results similar to radiologists.

A family of locally constrained CCA models for detecting activation patterns in fMRI

Canonical correlation analysis (CCA) has been used in Functional Magnetic Resonance Imaging (fMRI) for improved detection of activation by incorporating time series from multiple voxels in a local neighborhood. To improve the specificity of local CCA methods, spatial constraints were previously proposed. In this study, constraints are generalized by introducing a family model of spatial constraints for CCA to further increase both sensitivity and specificity in fMRI activation detection. The proposed locally-constrained CCA (cCCA) model is formulated in terms of a multivariate constrained optimization problem and solved efficiently with numerical optimization techniques. To evaluate the performance of this cCCA model, simulated data are generated with a Signal-To-Noise Ratio of 0.25, which is realistic to the noise level contained in episodic memory fMRI data. Receiver operating characteristic (ROC) methods are used to compare the performance of different models. The cCCA model with optimum parameters (called optimum-cCCA) obtains the largest area under the ROC curve. Furthermore, a novel validation method is proposed to validate the selected optimum-cCCA parameters based on ROC from simulated data and real fMRI data. Results for optimum-cCCA are then compared with conventional fMRI analysis methods using data from an episodic memory task. Wavelet-resampled resting-state data are used to obtain the null distribution of activation. For simulated data, accuracy in detecting activation increases for the optimum-cCCA model by about 43% as compared to the single voxel analysis with comparable Gaussian smoothing. Results from the real fMRI data set indicate a significant increase in activation detection, particularly in hippocampus, para-hippocampal area and nearby medial temporal lobe regions with the proposed method.

Comparison between diffusion-weighted imaging, T2-weighted, and postcontrast T1-weighted imaging after MR-guided, high intensity, focused ultrasound treatment of uterine leiomyomata: preliminary results

PURPOSE

To investigate the comparison between diffusion-weighted imaging (DWI), T2-weighted imaging, (T2WI) and contrast T1-weighted imaging (cT1WI) in uterine leiomyoma following treatment by magnetic resonance imaging-guided, high intensity focused ultrasound surgery (MRg-HIFUS).

METHODS

Twenty one patients (45 +/- 5 yrs) with clinical symptoms of uterine leiomyoma (fibroids) were treated by MRg-HIFUS using an integrated 1.5T MRI-HIFUS system. MRI parameters consisted of DWI, T2WI, and T1-weighted fast spoiled gradient echo before and after contrast. The post-MRg-HIFUS treatment volume in the fibroid was assessed by cT1WI and DWI. Trace apparent diffusion coefficient maps were constructed for quantitative analysis. The regions of the treated uterine tissue were defined by a semisupervised segmentation method called the "eigenimage filter," using both cT1WI and DWI. Signal-to-noise ratios were determined for the T2WI pretreatment images. Segmented regions were tested by a similarity index for congruence. Descriptive, regression, and Bland-Altman statistics were calculated.

RESULTS

All the patients exhibited heterogeneously increased DWI signal intensity localized in the treated fibroid regions and were colocalized with the cT1WI defined area. The mean pretreatment T2WI signal intensity ratios were T2WI/muscle = 1.8 +/- 0.7 and T2WI/myometrium = 0.7 +/- 0.4. The congruence between the regions was significant, with a similarity of 84% and a difference of 8% between the regions. Regression analyses of the cT1WI and DWI segmented treatment volume were found to be significantly correlated (r2 = 0.94, p < 0.05) with the linear equation, (cT1WI) = 1.1 (DWI)-0.66. There is good agreement between the regions defined by cT1WI and DWI in most of the cases as shown from the Bland-Altman plots.

CONCLUSIONS

Diffusion-weighted imaging exhibited excellent agreement, congruence, and correlation with the cT1WI-defined region of treatment in uterine fibroid. Therefore, DWI could be useful as an adjunct for assessing treatment of uterine fibroids by MRg-HIFUS.

Statistical Hypothesis Testing for Postreconstructed and Postregistered Medical Images

Postreconstructed and postregistered medical images are typically treated as the raw data, implicitly assuming that those operations are error free. We question this assumption and explore how the precision of reconstruction and affine registration can be assessed by the image covariance matrix and confidence interval, called the confidence eigenimage, using a statistical model-based approach. Various hypotheses may be tested after image reconstruction and registration using classical statistical hypothesis testing vehicles: Is there a statistically significant difference between images? Does the intensity at a specific location or area of interest belong to the "normal" range? Is there a tumor? Does the image require rigid registration? We illustrate statistical hypothesis testing with three examples: breast computed tomography, breast near infrared linear reconstruction, and brain magnetic resonance imaging.

Detection of spectral signatures in multispectral MR images for classification

This paper presents a new spectral signature detection approach to magnetic resonance (MR) image classification. It is called constrained energy minimization (CEM) method, which is derived from the minimum variance distortionless response in passive sensor array processing. It considers a bank of spectral channels as an array of sensors where each spectral channel represents a sensor and object spectral signature in multispectral MR images are viewed as signals impinging upon the array. The strength of the CEM lies on its ability in detection of spectral signatures of interest without knowing image background. The detected spectral signatures are then used for classification. The CEM makes use of a finite impulse response (FIR) filter to linearly constrain a desired object while minimizing interfering effects caused by other unknown signal sources. Unlike most spatial-based classification techniques, the proposed CEM takes advantage of spectral characteristics to achieve object detection and classification. A series of experiments is conducted and compared with the commonly used c-means method for performance evaluation. The results show that the CEM method is a promising and effective spectral technique for MR image classification.

Orthogonal subspace projection-based approaches to classification of MR image sequences

Orthogonal subspace projection (OSP) approach has shown success in hyperspectral image classification. Recently, the feasibility of applying OSP to multispectral image classification was also demonstrated via SPOT (Satellite Pour 1'Observation de la Terra) and Landsat (Land Satellite) images. Since an MR (magnetic resonance) image sequence is also acquired by multiple spectral channels (bands), this paper presents a new application of OSP in MR image classification. The idea is to model an MR image pixel in the sequence as a linear mixture of substances (such as white matter, gray matter, cerebral spinal fluid) of interest from which each of these substances can be classified by a specific subspace projection operator followed by a desired matched filter. The experimental results show that OSP provides a promising alternative to existing MR image classification techniques.

Identification of cerebral ischemic lesions in rat using Eigenimage filtered magnetic resonance imaging

An accurate noninvasive time-independent identification of an ischemic cerebral lesion is an important objective of magnetic resonance imaging (MRI). This study describes a novel application of a multiparameter MRI analysis algorithm, the Eigenimage (EI) filter, to experimental stroke. The EI is a linear filter that maximizes the projection of a desired tissue (ischemic tissue) while it minimizes the projection of undesired tissues (nonischemic tissue) onto a composite image called an eigenimage. Rats (n=26) were subjected to permanent middle cerebral artery occlusion. T2- and T1-weighted coronal MRI were acquired on separate groups of animals. The animals were immediately sacrificed after each imaging session for histopathological analysis of tissue at 4-8 h, 16-24 h, and 48-168 h after stroke onset. Lesion areas from MRI were defined using EI. The EI defined lesion areas were coregistered and warped to the corresponding histopathological sections. The ischemic lesion as defined by EI exhibited ischemic cell damage ranging from scattered acute cell damage to pan necrosis. Ischemic cellular damage was not detected in homologous contralateral hemisphere regions. EI lesion areas overlaid on histopathological sections were significantly correlated (r=0.92, p<0.05) acutely, (r=0.98, p<0.05) subacutely, and (r=0.99, p<0.05) chronically. These data indicate that EI methodology can accurately segment ischemic damage after MCA occlusion from 4-168 h after stroke.

Registration and warping of magnetic resonance images to histological sections

We present a method for coregistration and warping of magnetic resonance images (MRI) to histological sections for comparison purposes. This methodology consists of a modified head and hat surface-based registration algorithm followed by a new automated warping approach using nonlinear thin plate splines to compensate for distortions between the data sets. To test the methodology, 15 male Wistar rats were subjected to focal cerebral ischemia via permanent occlusion of the middle cerebral artery. The MRI images were acquired in separate groups of animals at 16-24 h (n = 9) and 48-168 h (n = 6) postocclusion. After imaging, animals were immediately sacrificed and hematoxylin- and eosin-stained brain sections were obtained for histological analysis. The MRI was coregistered and warped to histological sections. The MRI lesion areas were defined using the Eigenimage (EI) filter technique. The EI is a linear filter that maximizes the projection of a desired tissue (ischemic tissue) while it minimizes the projection of undesired tissues (nonischemic tissue) onto a composite image called an EI. When using coregistration without warping the MRI lesion area demonstrated poor correlation (r = 0.55, p > 0.01) with a percent difference between the two lesion areas of 22.5% +/- 10.8%. After warping, the MRI and histology had significant correlation (r = 0.97, p < 0.01) and a decreased percent difference of 5.56% +/- 4.31%. This methodology is simple and robust for coregistration and warping of MRI to histological sections and can be utilized in many applications for comparison of MRI to histological data.

Optimal linear transformation for MRI feature extraction

This paper presents development and application of a feature extraction method for magnetic resonance imaging (MRI), without explicit calculation of tissue parameters. A three-dimensional (3-D) feature space representation of the data is generated in which normal tissues are clustered around prespecified target positions and abnormalities are clustered elsewhere. This is accomplished by a linear minimum mean square error transformation of categorical data to target positions. From the 3-D histogram (cluster plot) of the transformed data, clusters are identified and regions of interest (ROI's) for normal and abnormal tissues are defined. These ROI's are used to estimate signature (prototype) vectors for each tissue type which in turn are used to segment the MRI scene. The proposed feature space is compared to those generated by tissue-parameter-weighted images, principal component images, and angle images, demonstrating its superiority for feature extraction and scene segmentation. Its relationship with discriminant analysis is discussed. The method and its performance are illustrated using a computer simulation and MRI images of an egg phantom and a human brain.