Deep learning approaches for detection and removal of ghosting artifacts in MR spectroscopy

Spectro-ViT: A vision transformer model for GABA-edited MEGA-PRESS reconstruction using spectrograms

This study investigated the use of a Vision Transformer (ViT) for reconstructing GABA-edited Magnetic Resonance Spectroscopy (MRS) data from a reduced number of transients. Transients refer to the samples collected during an MRS acquisition by repeating the experiment to generate a signal of sufficient quality. Specifically, 80 transients were used instead of the typical 320 transients, aiming to reduce scan time. The 80 transients were pre-processed and converted into a spectrogram image representation using the Short-Time Fourier Transform (STFT). A pre-trained ViT, named Spectro-ViT, was fine-tuned and then tested using in-vivo GABA-edited MEGA-PRESS data. Its performance was compared against other pipelines in the literature using quantitative quality metrics and estimated metabolite concentration values, with the typical 320-transient scans serving as the reference for comparison. The Spectro-ViT model exhibited the best overall quality metrics among all other pipelines against which it was compared. The metabolite concentrations from Spectro-ViT's reconstructions for GABA+ achieved the best average R2 value of 0.67 and the best average Mean Absolute Percentage Error (MAPE) value of 9.68%, with no significant statistical differences found compared to the 320-transient reference. The code to reproduce this research is available at https://github.com/MICLab-Unicamp/Spectro-ViT.

Automatic motion artifact detection in electrodermal activity signals using 1D U-net architecture

We developed a method for automated detection of motion and noise artifacts (MNA) in electrodermal activity (EDA) signals, based on a one-dimensional U-Net architecture. EDA has been widely employed in diverse applications to assess sympathetic functions. However, EDA signals can be easily corrupted by MNA, which frequently occur in wearable systems, particularly those used for ambulatory recording. MNA can lead to false decisions, resulting in inaccurate assessment and diagnosis. Several approaches have been proposed for MNA detection; however, questions remain regarding the generalizability and the feasibility of implementation of the algorithms in real-time especially those involving deep learning approaches. In this work, we propose a deep learning approach based on a one-dimensional U-Net architecture using spectrograms of EDA for MNA detection. We developed our method using four distinct datasets, including two independent testing datasets, with a total of 9602 128-s EDA segments from 104 subjects. Our proposed scheme, including data augmentation, spectrogram computation, and 1D U-Net, yielded balanced accuracies of 80.0 ± 13.7 % and 75.0 ± 14.0 % for the two independent test datasets; these results are better than or comparable to those of other five state-of-the-art methods. Additionally, the computation time of our feature computation and machine learning classification was significantly lower than that of other methods (p < .001). The model requires only 0.28 MB of memory, which is far smaller than the two deep learning approaches (4.93 and 54.59 MB) which were used as comparisons to our study. Our model can be implemented in real-time in embedded systems, even with limited memory and an inefficient microprocessor, without compromising the accuracy of MNA detection.

Two-Dimensional Laplace NMR Reconstruction through Deep Learning Enhancement

Laplace NMR is a powerful tool for studying molecular dynamics and spin interactions, providing diffusion and relaxation information that complements Fourier NMR used for composition determination and structure elucidation. However, Laplace NMR demands sophisticated signal processing algorithms such as inverse Laplace transform (ILT). Due to the inherently ill-posed nature of ILT problems, it is generally challenging to perform satisfactory Laplace NMR processing and reconstruction, particularly for two-dimensional Laplace NMR. Herein, we propose a proof-of-concept approach that blends a physics-informed strategy with data-driven deep learning for two-dimensional Laplace NMR reconstruction. This approach integrates prior knowledge of mathematical and physical laws governing multidimensional decay signals by constructing a forward process model to simulate relationships among different decay factors. Benefiting from a noniterative neural network algorithm that automatically acquires prior information from synthetic data during training, this approach avoids tedious parameter tuning and enhances user friendliness. Experimental results demonstrate the practical effectiveness of this approach. As an advanced and impactful technique, this approach brings a fresh perspective to multidimensional Laplace NMR inversion.

Comparison of convolutional-neural-networks-based method and LCModel on the quantification of in vivo magnetic resonance spectroscopy

BACKGROUND

Quantification of metabolites concentrations in institutional unit (IU) is important for inter-subject and long-term comparisons in the applications of magnetic resonance spectroscopy (MRS). Recently, deep learning (DL) algorithms have found a variety of applications on the process of MRS data. A quantification strategy compatible to DL base MRS spectral processing method is, therefore, useful.

MATERIALS AND METHODS

This study aims to investigate whether metabolite concentrations quantified using a convolutional neural network (CNN) based method, coupled with a scaling procedure that normalizes spectral signals for CNN input and linear regression, can effectively reflect variations in metabolite concentrations in IU across different brain regions with varying signal-to-noise ratios (SNR) and linewidths (LW). An error index based on standard error (SE) is proposed to indicate the confidence levels associated with metabolite predictions. In vivo MRS spectra were acquired from three brain regions of 43 subjects using a 3T system.

RESULTS

The metabolite concentrations in IU of five major metabolites, quantified using CNN and LCModel, exhibit similar ranges with Pearson's correlation coefficients ranging from 0.24 to 0.78. The SE of the metabolites shows a positive correlation with Cramer-Rao lower bound (CRLB) (r=0.46) and  absolute CRLB (r=0.81), calculated by multiplying CRLBs with the quantified metabolite content.

CONCLUSION

In conclusion, the CNN based method with the proposed scaling procedures can be employed to quantify in vivo MRS spectra and derive metabolites concentrations in IU. The SE can be used as error index, indicating predicted uncertainties for metabolites and sharing information similar to the absolute CRLB.

Results of the 2023 ISBI challenge to reduce GABA-edited MRS acquisition time

PURPOSE

Use a conference challenge format to compare machine learning-based gamma-aminobutyric acid (GABA)-edited magnetic resonance spectroscopy (MRS) reconstruction models using one-quarter of the transients typically acquired during a complete scan.

METHODS

There were three tracks: Track 1: simulated data, Track 2: identical acquisition parameters with in vivo data, and Track 3: different acquisition parameters with in vivo data. The mean squared error, signal-to-noise ratio, linewidth, and a proposed shape score metric were used to quantify model performance. Challenge organizers provided open access to a baseline model, simulated noise-free data, guides for adding synthetic noise, and in vivo data.

RESULTS

Three submissions were compared. A covariance matrix convolutional neural network model was most successful for Track 1. A vision transformer model operating on a spectrogram data representation was most successful for Tracks 2 and 3. Deep learning (DL) reconstructions with 80 transients achieved equivalent or better SNR, linewidth and fit error compared to conventional 320 transient reconstructions. However, some DL models optimized linewidth and SNR without actually improving overall spectral quality, indicating a need for more robust metrics.

CONCLUSION

DL-based reconstruction pipelines have the promise to reduce the number of transients required for GABA-edited MRS.

Fusion Networks of CNN and Transformer with Channel Attention for Accurate Tumor Imaging in Magnetic Particle Imaging

BACKGROUND

Magnetic Particle Imaging (MPI) is an emerging molecular imaging technique. However, since X-space reconstruction ignores system properties, it can lead to blurring of the reconstructed image, posing challenges for accurate quantification. To address this issue, we propose the use of deep learning to remove the blurry artifacts; (2) Methods: Our network architecture consists of a combination of Convolutional Neural Network (CNN) and Transformer. The CNN utilizes convolutional layers to automatically extract pixel-level local features and reduces the size of feature maps through pooling layers, effectively capturing local information within the images. The Transformer module is responsible for extracting contextual features from the images and efficiently capturing long-range dependencies, enabling a more effective modeling of global features in the images. By combining the features extracted by both CNN and Transformer, we capture both global and local features simultaneously, thereby improving the quality of reconstructed images; (3) Results: Experimental results demonstrate that the network effectively removes blurry artifacts from the images, and it exhibits high accuracy in precise tumor quantification. The proposed method shows superior performance over the state-of-the-art methods; (4) Conclusions: This bears significant implications for the image quality improvement and clinical application of MPI technology.

Denoising single MR spectra by deep learning: Miracle or mirage?

PURPOSE

The inherently poor SNR of MRS measurements presents a significant hurdle to its clinical application. Denoising by machine or deep learning (DL) was proposed as a remedy. It is investigated whether such denoising leads to lower estimate uncertainties or whether it essentially reduces noise in signal-free areas only.

METHODS

Noise removal based on supervised DL with U-nets was implemented using simulated 1 H MR spectra of human brain in two approaches: (1) via time-frequency domain spectrograms and (2) using 1D spectra as input. Quality of denoising was evaluated in three ways: (1) by an adapted fit quality score, (2) by traditional model fitting, and (3) by quantification via neural networks.

RESULTS

Visually appealing spectra were obtained; hinting that denoising is well-suited for MRS. However, an adapted denoising score showed that noise removal is inhomogeneous and more efficient for signal-free areas. This was confirmed by quantitative analysis of traditional fit results as well as DL quantitation following DL denoising. DL denoising, although apparently successful as judged by mean squared errors, led to substantially biased estimates in both implementations.

CONCLUSION

The implemented DL-based denoising techniques may be useful for display purposes, but do not help quantitative evaluations, confirming expectations based on estimation theory: Cramér Rao lower bounds defined by the original data and the appropriate fitting model cannot be circumvented in an unbiased way for single data sets, unless additional prior knowledge can be incurred in the form of parameter restrictions/relations or applicable substates.

Quantification of short echo time MRS signals with improved version of QUantitation based on quantum ESTimation algorithm

Magnetic resonance spectroscopy offers information about metabolite changes in the organism, which can be used in diagnosis. While short echo time proton spectra exhibit more distinguishable metabolites compared with proton spectra acquired with long echo times, their quantification (and providing estimates of metabolite concentrations) is more challenging. They are hampered by a background signal, which originates mainly from macromolecules (MM) and mobile lipids. An improved version of the quantification algorithm QUantitation based on quantum ESTimation (QUEST), with MM prior knowledge (QUEST-MM), dedicated to proton signals and invoking appropriate prior knowledge on MM, is proposed and tested. From a single acquisition, it enables better metabolite quantification, automatic estimation of the background, and additional automatic quantification of MM components, thus improving its applicability in the clinic. The proposed algorithm may facilitate studies that involve patients with pathological MM in the brain. QUEST-MM and three QUEST-based strategies for quantifying short echo time signals are compared in terms of bias-variance trade-off and Cramér-Rao lower bound estimates. The performances of the methods are evaluated through extensive Monte Carlo studies. In particular, the histograms of the metabolite and MM amplitude distributions demonstrate the performances of the estimators. They showed that QUEST-MM works better than QUEST (Subtract approach) and is a good alternative to QUEST when measured MM signal is unavailable or unsuitable. Quantification with QUEST-MM is shown for 1 H in vivo rat brain signals obtained with the SPECIAL pulse sequence at 9.4 T, and human brain signals obtained, respectively, with STEAM at 4 T and PRESS at 3 T. QUEST-MM is implemented in jMRUI and will be available for public use from version 7.1.

Quantification of spatially localized MRS by a novel deep learning approach without spectral fitting

PURPOSE

To propose a novel end-to-end deep learning model to quantify absolute metabolite concentrations from in vivo J-point resolved spectroscopy (JPRESS) without using spectral fitting.

METHODS

A novel encoder-decoder-style neural network was created, which was trained to predict metabolite concentrations and individual component signals concurrently from 3T JPRESS data in the time domain. The training data set contained 100 000 samples created by spin-density simulations using experimentally used RF pulses. Concentrations, phase, frequencies, linewidths, and T2 relaxation times in the training data set were varied over a large range with uniform distributions. Random synthesized noise and extraneous signals were added to the data set. Two thousand validation samples were created similarly to the training data set but with mean concentrations close to in vivo values. An in vivo test was conducted with 20 samples acquired from the human brain.

RESULTS

Both validation and in vivo test results showed that the proposed model successfully predicted metabolite concentrations as well as individual metabolite signals without involving spectral fitting, while extraneous peaks or unregistered signals were filtered out. Compared with the short-TE spectral fitting by LCModel, the proposed method had the advantage that the undesired correlations between the estimated concentrations and noise levels and between metabolites were eliminated or substantially reduced.

CONCLUSION

The proposed method provides a working deep learning model that directly maps in vivo JPRESS data to metabolite concentrations. Because spectral fitting is not used, the trained model does not depend on the assumptions associated with parameter tuning when applied to in vivo data.

Application of a 1H Brain MRS Benchmark Dataset to Deep Learning for Out-of-Voxel Artifacts

Neural networks are potentially valuable for many of the challenges associated with MRS data. The purpose of this manuscript is to describe the AGNOSTIC dataset, which contains 259,200 synthetic 1H MRS examples for training and testing neural networks. AGNOSTIC was created using 270 basis sets that were simulated across 18 field strengths and 15 echo times. The synthetic examples were produced to resemble in vivo brain data with combinations of metabolite, macromolecule, residual water signals, and noise. To demonstrate the utility, we apply AGNOSTIC to train two Convolutional Neural Networks (CNNs) to address out-of-voxel (OOV) echoes. A Detection Network was trained to identify the point-wise presence of OOV echoes, providing proof of concept for real-time detection. A Prediction Network was trained to reconstruct OOV echoes, allowing subtraction during post-processing. Complex OOV signals were mixed into 85% of synthetic examples to train two separate CNNs for the detection and prediction of OOV signals. AGNOSTIC is available through Dryad and all Python 3 code is available through GitHub. The Detection network was shown to perform well, identifying 95% of OOV echoes. Traditional modeling of these detected OOV signals was evaluated and may prove to be an effective method during linear-combination modeling. The Prediction Network greatly reduces OOV echoes within FIDs and achieved a median log10 normed-MSE of -1.79, an improvement of almost two orders of magnitude.

Evaluation of deep learning models for quality control of MR spectra

Purpose

While 3D MR spectroscopic imaging (MRSI) provides valuable spatial metabolic information, one of the hurdles for clinical translation is its interpretation, with voxel-wise quality control (QC) as an essential and the most time-consuming step. This work evaluates the accuracy of machine learning (ML) models for automated QC filtering of individual spectra from 3D healthy control and patient datasets.

Methods

A total of 53 3D MRSI datasets from prior studies (30 neurological diseases, 13 brain tumors, and 10 healthy controls) were included in the study. Three ML models were evaluated: a random forest classifier (RF), a convolutional neural network (CNN), and an inception CNN (ICNN) along with two hybrid models: CNN + RF, ICNN + RF. QC labels used for training were determined manually through consensus of two MRSI experts. Normalized and cropped real-valued spectra was used as input. A cross-validation approach was used to separate datasets into training/validation/testing sets of aggregated voxels.

Results

All models achieved a minimum AUC of 0.964 and accuracy of 0.910. In datasets from neurological disease and controls, the CNN model produced the highest AUC (0.982), while the RF model achieved the highest AUC in patients with brain tumors (0.976). Within tumor lesions, which typically exhibit abnormal metabolism, the CNN AUC was 0.973 while that of the RF was 0.969. Data quality inference times were on the order of seconds for an entire 3D dataset, offering drastic time reduction compared to manual labeling.

Conclusion

ML methods accurately and rapidly performed automated QC. Results in tumors highlights the applicability to a variety of metabolic conditions.

Using Single-Voxel Magnetic Resonance Spectroscopy Data Acquired at 1.5T to Classify Multivoxel Data at 3T: A Proof-of-Concept Study

In vivo magnetic resonance spectroscopy (MRS) has two modalities, single-voxel (SV) and multivoxel (MV), in which one or more contiguous grids of SVs are acquired.

PURPOSE

To test whether MV grids can be classified with models trained with SV.

METHODS

Retrospective study. Training dataset: Multicenter multiformat SV INTERPRET, 1.5T. Testing dataset: MV eTumour, 3T. Two classification tasks were completed: 3-class (meningioma vs. aggressive vs. normal) and 4-class (meningioma vs. low-grade glioma vs. aggressive vs. normal). Five different methods were tested for feature selection. The classification was implemented using linear discriminant analysis (LDA), random forest, and support vector machines. The evaluation was completed with balanced error rate (BER) and area under the curve (AUC) on both sets. The accuracy in class prediction was calculated by developing a solid tumor index (STI) and segmentation accuracy with the Dice score.

RESULTS

The best method was sequential forward feature selection combined with LDA, with AUCs = 0.95 (meningioma), 0.89 (aggressive), 0.82 (low-grade glioma), and 0.82 (normal). STI was 66% (4-class task) and 71% (3-class task) because two cases failed completely and two more had suboptimal STI as defined by us.

DISCUSSION

The reasons for failure in the classification of the MV test set were related to the presence of artifacts.

A review of machine learning applications for the proton MR spectroscopy workflow

This literature review presents a comprehensive overview of machine learning (ML) applications in proton MR spectroscopy (MRS). As the use of ML techniques in MRS continues to grow, this review aims to provide the MRS community with a structured overview of the state-of-the-art methods. Specifically, we examine and summarize studies published between 2017 and 2023 from major journals in the MR field. We categorize these studies based on a typical MRS workflow, including data acquisition, processing, analysis, and artificial data generation. Our review reveals that ML in MRS is still in its early stages, with a primary focus on processing and analysis techniques, and less attention given to data acquisition. We also found that many studies use similar model architectures, with little comparison to alternative architectures. Additionally, the generation of artificial data is a crucial topic, with no consistent method for its generation. Furthermore, many studies demonstrate that artificial data suffers from generalization issues when tested on in vivo data. We also conclude that risks related to ML models should be addressed, particularly for clinical applications. Therefore, output uncertainty measures and model biases are critical to investigate. Nonetheless, the rapid development of ML in MRS and the promising results from the reviewed studies justify further research in this field.

Quantification of MR spectra by deep learning in an idealized setting: Investigation of forms of input, network architectures, optimization by ensembles of networks, and training bias

PURPOSE

The aims of this work are (1) to explore deep learning (DL) architectures, spectroscopic input types, and learning designs toward optimal quantification in MR spectroscopy of simulated pathological spectra; and (2) to demonstrate accuracy and precision of DL predictions in view of inherent bias toward the training distribution.

METHODS

Simulated 1D spectra and 2D spectrograms that mimic an extensive range of pathological in vivo conditions are used to train and test 24 different DL architectures. Active learning through altered training and testing data distributions is probed to optimize quantification performance. Ensembles of networks are explored to improve DL robustness and reduce the variance of estimates. A set of scores compares performances of DL predictions and traditional model fitting (MF).

RESULTS

Ensembles of heterogeneous networks that combine 1D frequency-domain and 2D time-frequency domain spectrograms as input perform best. Dataset augmentation with active learning can improve performance, but gains are limited. MF is more accurate, although DL appears to be more precise at low SNR. However, this overall improved precision originates from a strong bias for cases with high uncertainty toward the dataset the network has been trained with, tending toward its average value.

CONCLUSION

MF mostly performs better compared to the faster DL approach. Potential intrinsic biases on training sets are dangerous in a clinical context that requires the algorithm to be unbiased to outliers (i.e., pathological data). Active learning and ensemble of networks are good strategies to improve prediction performances. However, data quality (sufficient SNR) has proven as a bottleneck for adequate unbiased performance-like in the case of MF.

Superconducting magnet designs and MRI accessibility: A review

Presently, magnetic resonance imaging (MRI) magnets must deliver excellent magnetic field (B₀ ) uniformity to achieve optimum image quality. Long magnets can satisfy the homogeneity requirements but require considerable superconducting material. These designs result in large, heavy, and costly systems that aggravate as field strength increases. Furthermore, the tight temperature tolerance of niobium titanium magnets adds instability to the system and requires operation at liquid helium temperature. These issues are crucial factors in the disparity of MR density and field strength use across the globe. Low-income settings show reduced access to MRI, especially to high field strengths. This article summarizes the proposed modifications to MRI superconducting magnet design and their impact on accessibility, including compact, reduced liquid helium, and specialty systems. Reducing the amount of superconductor inevitably entails shrinking the magnet size, resulting in higher field inhomogeneity. This work also reviews the state-of-the-art imaging and reconstruction methods to overcome this issue. Finally, we summarize the current and future challenges and opportunities in the design of accessible MRI.

Impact of gradient scheme and non-linear shimming on out-of-voxel echo artifacts in edited MRS

Out-of-voxel (OOV) signals are common spurious echo artifacts in MRS. These signals often manifest in the spectrum as very strong "ripples," which interfere with spectral quantification by overlapping with targeted metabolite resonances. Dephasing optimization through coherence order pathway selection (DOTCOPS) gradient schemes are algorithmically optimized to suppress all potential alternative coherence transfer pathways (CTPs), and should suppress unwanted OOV echoes. In addition, second-order shimming uses non-linear gradient fields to maximize field homogeneity inside the voxel, which unfortunately increases the diversity of local gradient fields outside of the voxel. Given that strong local spatial B0 gradients can refocus unintended CTPs, it is possible that OOVs are less prevalent when only linear first-order shimming is applied. Here we compare the size of unwanted OOV signals in Hadamard-edited (HERMES) data acquired with either a local gradient scheme (which we refer to here as "Shared") or DOTCOPS, and with first- or second-order shimming. We collected data from 15 healthy volunteers in two brain regions (voxel size 30 × 26 × 26 mm3 ) from which it is challenging to acquire MRS data: medial prefrontal cortex and left temporal cortex. Characteristic OOV echoes were seen in both GABA- and GSH-edited spectra for both brain regions, gradient schemes, and shimming approaches. A linear mixed-effect model revealed a statistically significant difference in the average residual based on the gradient scheme in both GABA- (p < 0.001) and GSH-edited (p < 0.001) spectra: that is, the DOTCOPS gradient scheme resulted in smaller OOV artifacts compared with the Shared scheme. There were no significant differences in OOV artifacts associated with shimming method. Thus, these results suggest that the DOTCOPS gradient scheme for J-difference-edited PRESS acquisitions yields spectra with smaller OOV echo artifacts than the Shared gradient scheme implemented in a widely disseminated editing sequence.

In vivo magnetic resonance 31 P-Spectral Analysis With Neural Networks: 31P-SPAWNN

PURPOSE

We have introduced an artificial intelligence framework, 31P-SPAWNN, in order to fully analyze phosphorus-31 (31 $$ {}^{31} $$ P) magnetic resonance spectra. The flexibility and speed of the technique rival traditional least-square fitting methods, with the performance of the two approaches, are compared in this work.

THEORY AND METHODS

Convolutional neural network architectures have been proposed for the analysis and quantification of31 $$ {}^{31} $$ P-spectroscopy. The generation of training and test data using a fully parameterized model is presented herein. In vivo unlocalized free induction decay and three-dimensional31 $$ {}^{31} $$ P-magnetic resonance spectroscopy imaging data were acquired from healthy volunteers before being quantified using either 31P-SPAWNN or traditional least-square fitting techniques.

RESULTS

The presented experiment has demonstrated both the reliability and accuracy of 31P-SPAWNN for estimating metabolite concentrations and spectral parameters. Simulated test data showed improved quantification using 31P-SPAWNN compared with LCModel. In vivo data analysis revealed higher accuracy at low signal-to-noise ratio using 31P-SPAWNN, yet with equivalent precision. Processing time using 31P-SPAWNN can be further shortened up to two orders of magnitude.

CONCLUSION

The accuracy, reliability, and computational speed of the method open new perspectives for integrating these applications in a clinical setting.

Convolutional neural networks to predict brain tumor grades and Alzheimer’s disease with MR spectroscopic imaging data

Purpose

To evaluate the value of convolutional neural network (CNN) in the diagnosis of human brain tumor or Alzheimer’s disease by MR spectroscopic imaging (MRSI) and to compare its Matthews correlation coefficient (MCC) score against that of other machine learning methods and previous evaluation of the same data. We address two challenges: 1) limited number of cases in MRSI datasets and 2) interpretability of results in the form of relevant spectral regions.

Methods

A shallow CNN with only one hidden layer and an ad-hoc loss function was constructed involving two branches for processing spectral and image features of a brain voxel respectively. Each branch consists of a single convolutional hidden layer. The output of the two convolutional layers is merged and fed to a classification layer that outputs class predictions for the given brain voxel.

Results

Our CNN method separated glioma grades 3 and 4 and identified Alzheimer’s disease patients using MRSI and complementary MRI data with high MCC score (Area Under the Curve were 0.87 and 0.91 respectively). The results demonstrated superior effectiveness over other popular methods as Partial Least Squares or Support Vector Machines. Also, our method automatically identified the spectral regions most important in the diagnosis process and we show that these are in good agreement with existing biomarkers from the literature.

Conclusion

Shallow CNNs models integrating image and spectral features improved quantitative and exploration and diagnosis of brain diseases for research and clinical purposes. Software is available at https://bitbucket.org/TeslaH2O/cnn_mrsi.

Developments in proton MR spectroscopic imaging of prostate cancer

In this paper, we review the developments of 1H-MR spectroscopic imaging (MRSI) methods designed to investigate prostate cancer, covering key aspects such as specific hardware, dedicated pulse sequences for data acquisition and data processing and quantification techniques. Emphasis is given to recent advancements in MRSI methodologies, as well as future developments, which can lead to overcome difficulties associated with commonly employed MRSI approaches applied in clinical routine. This includes the replacement of standard PRESS sequences for volume selection, which we identified as inadequate for clinical applications, by sLASER sequences and implementation of 1H MRSI without water signal suppression. These may enable a new evaluation of the complementary role and significance of MRSI in prostate cancer management.

Improved signal-to-noise performance of MultiNet GRAPPA 1 H FID MRSI reconstruction with semi-synthetic calibration data

PURPOSE

To further develop MultiNet GRAPPA, a neural-network-based reconstruction, for lower SNR proton MRSI (¹ H MRSI) data using adapted undersampling schemes and improved training sets.

METHODS

¹ H FID-MRSI data and an anatomical image for GRAPPA reconstruction were acquired in two slices in the human brain (n = 6) at 7T. MRSI data were retrospectively undersampled for a 4×, 6×, and 7× acceleration rate. Signal-to-noise, relative error (RE) between accelerated and fully sampled metabolic maps, RMS of the lipid artifacts, and fitting reliability were compared across acceleration rates, to the fully sampled data, and with different kinds and amounts of training images.

RESULTS

Training with semi-synthetic images resulted in higher SNR and lower lipid RMS relative to training with acquired images from one or several subjects. SNR increased with the number of semi-synthetic training images and the 4× accelerated data retains ∼30% more SNR than other accelerated data. Spectra reconstructed with 20 semi-synthetic averages retained ∼100% more SNR and had ∼5% lower lipid RMS than those reconstructed with the center k-space points of one image as was originally proposed for very high SNR MRSI data and had higher fitting reliability. The metabolite RE was lowest when training with 20-semi-synthetic training images and highest when training with the center k-space points of one image.

CONCLUSION

MultiNet GRAPPA is feasible with lower SNR ¹ H MRSI data if 20-semi-synthetic training images are used at a 4× acceleration rate. This acceleration rate provided the best trade-off between scan time and spectral SNR.

Extraction of artefactual MRS patterns from a large database using non-negative matrix factorization

Despite the success of automated pattern recognition methods in problems of human brain tumor diagnostic classification, limited attention has been paid to the issue of automated data quality assessment in the field of MRS for neuro-oncology. Beyond some early attempts to address this issue, the current standard in practice is MRS quality control through human (expert-based) assessment. One aspect of automatic quality control is the problem of detecting artefacts in MRS data. Artefacts, whose variety has already been reviewed in some detail and some of which may even escape human quality control, have a negative influence in pattern recognition methods attempting to assist tumor characterization. The automatic detection of MRS artefacts should be beneficial for radiology as it guarantees more reliable tumor characterizations, as well as the development of more robust pattern recognition-based tumor classifiers and more trustable MRS data processing and analysis pipelines. Feature extraction methods have previously been used to help distinguishing between good and bad quality spectra to apply subsequent supervised pattern recognition techniques. In this study, we apply feature extraction differently and use a variant of a method for blind source separation, namely Convex Non-Negative Matrix Factorization, to unveil MRS signal sources in a completely unsupervised way. We hypothesize that, while most sources will correspond to the different tumor patterns, some of them will reflect signal artefacts. The experimental work reported in this paper, analyzing a combined short and long echo time ¹ H-MRS database of more than 2000 spectra acquired at 1.5T and corresponding to different tumor types and other anomalous masses, provides a first proof of concept that points to the possible validity of this approach.

Identifying the source of spurious signals caused by B0 inhomogeneities in single-voxel 1 H MRS

Purpose

Single‐voxel MRS (SV MRS) requires robust volume localization as well as optimized crusher and phase‐cycling schemes to reduce artifacts arising from signal outside the volume of interest. However, due to local magnetic field gradients (B₀ inhomogeneities), signal that was dephased by the crusher gradients during acquisition might rephase, leading to artifacts in the spectrum. Here, we analyzed this mechanism, aiming to identify the source of signals arising from unwanted coherence pathways (spurious signals) in SV MRS from a B₀ map.

Methods

We investigated all possible coherence pathways associated with imperfect localization in a semi‐localized by adiabatic selective refocusing (semi‐LASER) sequence for potential rephasing of signals arising from unwanted coherence pathways by a local magnetic field gradient. We searched for locations in the B₀ map where the signal dephasing due to external (crusher) and internal (B₀) field gradients canceled out. To confirm the mechanism, SV‐MR spectra (TE = 31 ms) and 3D‐CSI data with the same volume localization as the SV experiments were acquired from a phantom and 2 healthy volunteers.

Results

Our analysis revealed that potential sources of spurious signals were scattered over multiple locations throughout the brain. This was confirmed by 3D‐CSI data. Moreover, we showed that the number of potential locations where spurious signals could originate from monotonically decreases with crusher strength.

Conclusion

We proposed a method to identify the source of spurious signals in SV ¹H MRS using a B₀ map. This can facilitate MRS sequence design to be less sensitive to experimental artifacts.

Bayesian deep learning-based 1 H-MRS of the brain: Metabolite quantification with uncertainty estimation using Monte Carlo dropout

PURPOSE

To develop a Bayesian convolutional neural network (BCNN) with Monte Carlo dropout sampling for metabolite quantification with simultaneous uncertainty estimation in deep learning-based proton MRS of the brain.

METHODS

Human brain spectra were simulated using basis spectra for 17 metabolites and macromolecules (N = 100 000) at 3.0 Tesla. In addition, actual in vivo spectra (N = 5) were modified by adjusting SNR and linewidth with increasing severity of spectral degradation (N = 50). A BCNN was trained on the simulated spectra to generate a noise-free, line-narrowed, macromolecule signal-removed, metabolite-only spectrum from a typical human brain spectrum. At inference, each input spectrum was Monte Carlo dropout sampled (50 times), and the resulting mean spectrum and variance spectrum were used for metabolite quantification and uncertainty estimation, respectively.

RESULTS

Using the simulated spectra, the mean absolute percent errors of the BCNN-predicted metabolite content were < 10% for Cr, Glu, Gln, mI, NAA, and Tau (< 5% for Glu, NAA, and mI). For all metabolites, the correlations (r's) between the ground-truth error and BCNN-predicted uncertainty ranged 0.72-0.94 (0.83 ± 0.06; p < 0.001). Using the modified in vivo spectra, the extent of variation in the estimated metabolite content against the increasing severity of spectral degradation tended to be smaller with BCNN than with linear combination of model spectra (LCModel). Overall, the variation in metabolite content tended to be more highly correlated with the uncertainty from BCNN than with the Cramér-Rao lower-bounds from LCModel (0.938 ± 0.019 vs. 0.881 ± 0.057 [p = 0.115]).

CONCLUSION

The BCNN with Monte Carlo dropout sampling may be used in deep learning-based MRS for the estimation of uncertainty in the machine-predicted metabolite content, which is important in the clinical application of deep learning-based MRS.

ArtifactID: Identifying artifacts in low-field MRI of the brain using deep learning

Low-field MR scanners are more accessible in resource-constrained settings where skilled personnel are scarce. Images acquired in such scenarios are prone to artifacts such as wrap-around and Gibbs ringing. Such artifacts negatively affect the diagnostic quality and may be confused with pathology or reduce the region of interest visibility. As a first step solution, ArtifactID identifies wrap-around and Gibbs ringing in low-field brain MRI. We utilized two datasets: 179 T₁-weighted pathological brain images from a 0.36 T scanner and 581 publicly available T₁-weighted brain images. Individual binary classification models were trained to identify through-plane wrap-around, in-plane wrap-around, and Gibbs ringing. Visual explanations obtained via the GradCAM method helped develop trust in the wrap-around model. The mean precision and recall metrics across the four implemented models were 97.6% and 92.83% respectively. Agreement analysis of the models and the radiologists' labels returned Cohen's kappa values of 0.768 ± 0.062, 1.00 ± 0.000, 0.89 ± 0.085, and 0.878 ± 0.103 for the through-plane wrap-around, in-plane wrap-around, and Gibbs ringing models, respectively.

Your mileage may vary: impact of data input method for a deep learning bone age app's predictions

OBJECTIVE

The purpose of this study was to evaluate agreement in predictions made by a bone age prediction application ("app") among three data input methods.

METHODS

The 16Bit Bone Age app is a browser-based deep learning application for predicting bone age on pediatric hand radiographs; recommended data input methods are direct image file upload or smartphone-capture of image. We collected 50 hand radiographs, split equally among 5 bone age groups. Three observers used the 16Bit Bone Age app to assess these images using 3 different data input methods: (1) direct image upload, (2) smartphone photo of image in radiology reading room, and (3) smartphone photo of image in a clinic.

RESULTS

Interobserver agreement was excellent for direct upload (ICC = 1.00) and for photos in reading room (ICC = 0.96) and good for photos in clinic (ICC = 0.82), respectively. Intraobserver agreement for the entire test set across the 3 data input methods was variable with ICCs of 0.95, 0.96, and 0.57 for the 3 observers, respectively.

DISCUSSION

Our findings indicate that different data input methods can result in discordant bone age predictions from the 16Bit Bone Age app. Further study is needed to determine the impact of data input methods, such as smartphone image capture, on deep learning app performance and accuracy.

Quality control of 3D MRSI data in glioblastoma: Can we do without the experts?

PURPOSE

Proton magnetic resonance spectroscopic imaging (1H MRSI) is a noninvasive technique for assessing tumor metabolism. Manual inspection is still the gold standard for quality control (QC) of spectra, but it is both time-consuming and subjective. The aim of the present study was to assess automatic QC of glioblastoma MRSI data using random forest analysis.

METHODS

Data for 25 patients, acquired prospectively in a preradiotherapy examination, were submitted to postprocessing with syngo.MR Spectro (VB40A; Siemens) or Java-based magnetic resonance user interface (jMRUI) software. A total of 28 features were extracted from each spectrum for the automatic QC. Three spectroscopists also performed manual inspections, labeling each spectrum as good or poor quality. All statistical analyses, with addressing unbalanced data, were conducted with R 3.6.1 (R Foundation for Statistical Computing; https://www.r-project.org).

RESULTS

The random forest method classified the spectra with an area under the curve of 95.5%, sensitivity of 95.8%, and specificity of 81.7%. The most important feature for the classification was Residuum_Lipids_Versus_Fit, obtained with syngo.MR Spectro.

CONCLUSION

The automatic QC method was able to distinguish between good- and poor-quality spectra, and can be used by radiation oncologists who are not spectroscopy experts. This study revealed a novel set of MRSI signal features that are closely correlated with spectral quality.

Deep Learning and Its Application to Function Approximation for MR in Medicine: An Overview

This article presents an overview of deep learning (DL) and its applications to function approximation for MR in medicine. The aim of this article is to help readers develop various applications of DL. DL has made a large impact on the literature of many medical sciences, including MR. However, its technical details are not easily understandable for non-experts of machine learning (ML).The first part of this article presents an overview of DL and its related technologies, such as artificial intelligence (AI) and ML. AI is explained as a function that can receive many inputs and produce many outputs. ML is a process of fitting the function to training data. DL is a kind of ML, which uses a composite of many functions to approximate the function of interest. This composite function is called a deep neural network (DNN), and the functions composited into a DNN are called layers. This first part also covers the underlying technologies required for DL, such as loss functions, optimization, initialization, linear layers, non-linearities, normalization, recurrent neural networks, regularization, data augmentation, residual connections, autoencoders, generative adversarial networks, model and data sizes, and complex-valued neural networks.The second part of this article presents an overview of the applications of DL in MR and explains how functions represented as DNNs are applied to various applications, such as RF pulse, pulse sequence, reconstruction, motion correction, spectroscopy, parameter mapping, image synthesis, and segmentation.

Estimation and removal of spurious echo artifacts in single-voxel MRS using sensitivity encoding

PURPOSE

In localized MRS, spurious echo artifacts commonly occur when unsuppressed signal outside the volume of interest is excited and refocused. In the spectral domain, these signals often overlap with metabolite resonances and hinder accurate quantification. Because the artifacts originate from regions separate from the target MRS voxel, this work proposes that sensitivity encoding based on receive-coil sensitivity profiles may be used to separate these signal contributions.

METHODS

Numerical simulations were performed to explore the effect of sensitivity-encoded separation for unknown artifact regions. An imaging-based approach was developed to identify regions that may contribute to spurious echo artifacts, and tested for sensitivity-based unfolding of signal on six data sets from three brain regions. Spectral data reconstructed using the proposed method ("ERASE") were compared with the standard coil combination.

RESULTS

The method was able to fully unfold artifact signals if regions were known a priori. Mismatch between estimated and true artifact regions reduced the efficiency of removal, yet metabolite signals were unaffected. Water suppression imaging was able to identify regions of unsuppressed signal, and ERASE (from up to eight regions) led to visible removal of artifacts relative to standard reconstruction. Fitting errors across major metabolites were also lower; for example, Cramér-Rao lower bounds of myo-inositol were 13.7% versus 17.5% for ERASE versus standard reconstruction, respectively.

CONCLUSION

The ERASE reconstruction tool was demonstrated to reduce spurious echo artifacts in single-voxel MRS. This tool may be incorporated into standard workflows to improve spectral quality when hardware limitations or other factors result in out-of-voxel signal contamination.

Determination of the Geographical Origin of Coffee Beans Using Terahertz Spectroscopy Combined With Machine Learning Methods

Different geographical origins can lead to great variance in coffee quality, taste, and commercial value. Hence, controlling the authenticity of the origin of coffee beans is of great importance for producers and consumers worldwide. In this study, terahertz (THz) spectroscopy, combined with machine learning methods, was investigated as a fast and non-destructive method to classify the geographic origin of coffee beans, comparing it with the popular machine learning methods, including convolutional neural network (CNN), linear discriminant analysis (LDA), and support vector machine (SVM) to obtain the best model. The curse of dimensionality will cause some classification methods which are struggling to train effective models. Thus, principal component analysis (PCA) and genetic algorithm (GA) were applied for LDA and SVM to create a smaller set of features. The first nine principal components (PCs) with an accumulative contribution rate of 99.9% extracted by PCA and 21 variables selected by GA were the inputs of LDA and SVM models. The results demonstrate that the excellent classification (accuracy was 90% in a prediction set) could be achieved using a CNN method. The results also indicate variable selecting as an important step to create an accurate and robust discrimination model. The performances of LDA and SVM algorithms could be improved with spectral features extracted by PCA and GA. The GA-SVM has achieved 75% accuracy in a prediction set, while the SVM and PCA-SVM have achieved 50 and 65% accuracy, respectively. These results demonstrate that THz spectroscopy, together with machine learning methods, is an effective and satisfactory approach for classifying geographical origins of coffee beans, suggesting the techniques to tap the potential application of deep learning in the authenticity of agricultural products while expanding the application of THz spectroscopy.

Preprocessing, analysis and quantification in single-voxel magnetic resonance spectroscopy: experts' consensus recommendations

Once an MRS dataset has been acquired, several important steps must be taken to obtain the desired metabolite concentration measures. First, the data must be preprocessed to prepare them for analysis. Next, the intensity of the metabolite signal(s) of interest must be estimated. Finally, the measured metabolite signal intensities must be converted into scaled concentration units employing a quantitative reference signal to allow meaningful interpretation. In this paper, we review these three main steps in the post-acquisition workflow of a single-voxel MRS experiment (preprocessing, analysis and quantification) and provide recommendations for best practices at each step.

Advanced magnetic resonance spectroscopic neuroimaging: Experts' consensus recommendations

Magnetic resonance spectroscopic imaging (MRSI) offers considerable promise for monitoring metabolic alterations associated with disease or injury; however, to date, these methods have not had a significant impact on clinical care, and their use remains largely confined to the research community and a limited number of clinical sites. The MRSI methods currently implemented on clinical MRI instruments have remained essentially unchanged for two decades, with only incremental improvements in sequence implementation. During this time, a number of technological developments have taken place that have already greatly benefited the quality of MRSI measurements within the research community and which promise to bring advanced MRSI studies to the point where the technique becomes a true imaging modality, while making the traditional review of individual spectra a secondary requirement. Furthermore, the increasing use of biomedical MR spectroscopy studies has indicated clinical areas where advanced MRSI methods can provide valuable information for clinical care. In light of this rapidly changing technological environment and growing understanding of the value of MRSI studies for biomedical studies, this article presents a consensus from a group of experts in the field that reviews the state-of-the-art for clinical proton MRSI studies of the human brain, recommends minimal standards for further development of vendor-provided MRSI implementations, and identifies areas which need further technical development.

Deep learning can accelerate and quantify simulated localized correlated spectroscopy

Nuclear magnetic resonance spectroscopy (MRS) allows for the determination of atomic structures and concentrations of different chemicals in a biochemical sample of interest. MRS is used in vivo clinically to aid in the diagnosis of several pathologies that affect metabolic pathways in the body. Typically, this experiment produces a one dimensional (1D) 1H spectrum containing several peaks that are well associated with biochemicals, or metabolites. However, since many of these peaks overlap, distinguishing chemicals with similar atomic structures becomes much more challenging. One technique capable of overcoming this issue is the localized correlated spectroscopy (L-COSY) experiment, which acquires a second spectral dimension and spreads overlapping signal across this second dimension. Unfortunately, the acquisition of a two dimensional (2D) spectroscopy experiment is extremely time consuming. Furthermore, quantitation of a 2D spectrum is more complex. Recently, artificial intelligence has emerged in the field of medicine as a powerful force capable of diagnosing disease, aiding in treatment, and even predicting treatment outcome. In this study, we utilize deep learning to: (1) accelerate the L-COSY experiment and (2) quantify L-COSY spectra. All training and testing samples were produced using simulated metabolite spectra for chemicals found in the human body. We demonstrate that our deep learning model greatly outperforms compressed sensing based reconstruction of L-COSY spectra at higher acceleration factors. Specifically, at four-fold acceleration, our method has less than 5% normalized mean squared error, whereas compressed sensing yields 20% normalized mean squared error. We also show that at low SNR (25% noise compared to maximum signal), our deep learning model has less than 8% normalized mean squared error for quantitation of L-COSY spectra. These pilot simulation results appear promising and may help improve the efficiency and accuracy of L-COSY experiments in the future.

An anomaly detection approach to identify chronic brain infarcts on MRI

The performance of current machine learning methods to detect heterogeneous pathology is limited by the quantity and quality of pathology in medical images. A possible solution is anomaly detection; an approach that can detect all abnormalities by learning how 'normal' tissue looks like. In this work, we propose an anomaly detection method using a neural network architecture for the detection of chronic brain infarcts on brain MR images. The neural network was trained to learn the visual appearance of normal appearing brains of 697 patients. We evaluated its performance on the detection of chronic brain infarcts in 225 patients, which were previously labeled. Our proposed method detected 374 chronic brain infarcts (68% of the total amount of brain infarcts) which represented 97.5% of the total infarct volume. Additionally, 26 new brain infarcts were identified that were originally missed by the radiologist during radiological reading. Our proposed method also detected white matter hyperintensities, anomalous calcifications, and imaging artefacts. This work shows that anomaly detection is a powerful approach for the detection of multiple brain abnormalities, and can potentially be used to improve the radiological workflow efficiency by guiding radiologists to brain anomalies which otherwise remain unnoticed.

Unsupervised anomaly detection using generative adversarial networks in 1H-MRS of the brain

The applicability of generative adversarial networks (GANs) capable of unsupervised anomaly detection (AnoGAN) was investigated in the management of quality of ¹H-MRS human brain spectra at 3.0 T. The AnoGAN was trained in an unsupervised manner solely on simulated normal brain spectra and used for filtering out abnormal spectra with a broad range of abnormalities, which were simulated by including abnormal ranges of SNR, linewidth and metabolite concentrations and spectral artifacts such as ghost, residual water, and lipid. The AnoGAN was able to filter out those spectra with SNR less than ~11-12 dB with an accuracy of ~80% or higher (assuming a normal SNR range to be 15-18 dB). It also detected with an accuracy of ~80% or higher those spectra, in which NAA levels were reduced by ~25-30% or more from the lower bound and elevated by ~20-30% or more from the upper bound of the normal concentration range (7.5-17 mmol/L), while the concentrations of the rest of the metabolites were all within the normal ranges. Despite the fact that those spectra contaminated with ghost, residual water or lipid have never been involved in the training or optimization of the AnoGAN, they were correctly classified as abnormal regardless of the types of the artifacts, depending solely on their intensity. Although the current version of our AnoGAN requires further technical improvement particularly for the detection of linewidth-associated abnormality and validation on in vivo data, our unsupervised deep learning-based approach could be an option in addition to those previously reported supervised deep learning-based approaches in the binary classification of spectral quality with an extended abnormal spectra regime.

The clinical utility of proton magnetic resonance spectroscopy in traumatic brain injury: recommendations from the ENIGMA MRS working group

Proton (1H) magnetic resonance spectroscopy provides a non-invasive and quantitative measure of brain metabolites. Traumatic brain injury impacts cerebral metabolism and a number of research groups have successfully used this technique as a biomarker of injury and/or outcome in both pediatric and adult TBI populations. However, this technique is underutilized, with studies being performed primarily at centers with access to MR research support. In this paper we present a technical introduction to the acquisition and analysis of in vivo 1H magnetic resonance spectroscopy and review 1H magnetic resonance spectroscopy findings in different injury populations. In addition, we propose a basic 1H magnetic resonance spectroscopy data acquisition scheme (Supplemental Information) that can be added to any imaging protocol, regardless of clinical magnetic resonance platform. We outline a number of considerations for study design as a way of encouraging the use of 1H magnetic resonance spectroscopy in the study of traumatic brain injury, as well as recommendations to improve data harmonization across groups already using this technique.

MC2 -Net: motion correction network for multi-contrast brain MRI

PURPOSE

A motion-correction network for multi-contrast brain MRI is proposed to correct in-plane rigid motion artifacts in brain MR images using deep learning.

METHOD

The proposed method consists of 2 parts: image alignment and motion correction. Alignment of multi-contrast MR images is performed in an unsupervised manner by a CNN work, yielding transformation parameters to align input images in order to minimize the normalized cross-correlation loss among multi-contrast images. Then, fine-tuning for image alignment is performed by maximizing the normalized mutual information. The motion correction network corrects motion artifacts in the aligned multi-contrast images. The correction network is trained to minimize the structural similarity loss and the VGG loss in a supervised manner. All datasets of motion-corrupted images are generated using motion simulation based on MR physics.

RESULTS

A motion-correction network for multi-contrast brain MRI successfully corrected artifacts of simulated motion for 4 test subjects, showing 0.96%, 7.63%, and 5.03% increases in the average structural simularity and 5.19%, 10.2%, and 7.48% increases in the average normalized mutual information for T₁ -weighted, T₂ -weighted, and T₂ -weighted fluid-attenuated inversion recovery images, respectively. The experimental setting with image alignment and artifact-free input images for other contrasts shows better performances in correction of simulated motion artifacts. Furthermore, the proposed method quantitatively outperforms recent deep learning motion correction and synthesis methods. Real motion experiments from 5 healthy subjects demonstrate the potential of the proposed method for use in a clinical environment.

CONCLUSION

A deep learning-based motion correction method for multi-contrast MRI was successfully developed, and experimental results demonstrate the validity of the proposed method.

Frequency and phase correction of J-difference edited MR spectra using deep learning

PURPOSE

To investigate whether a deep learning-based (DL) approach can be used for frequency-and-phase correction (FPC) of MEGA-edited MRS data.

METHODS

Two neural networks (1 for frequency, 1 for phase) consisting of fully connected layers were trained and validated using simulated MEGA-edited MRS data. This DL-FPC was subsequently tested and compared to a conventional approach (spectral registration [SR]) and to a model-based SR implementation (mSR) using in vivo MEGA-edited MRS datasets. Additional artificial offsets were added to these datasets to further investigate performance.

RESULTS

The validation showed that DL-based FPC was capable of correcting within 0.03 Hz of frequency and 0.4°of phase offset for unseen simulated data. DL-based FPC performed similarly to SR for the unmanipulated in vivo test datasets. When additional offsets were added to these datasets, the networks still performed well. However, although SR accurately corrected for smaller offsets, it often failed for larger offsets. The mSR algorithm performed well for larger offsets, which was because the model was generated from the in vivo datasets. In addition, the computation times were much shorter using DL-based FPC or mSR compared to SR for heavily distorted spectra.

CONCLUSION

These results represent a proof of principle for the use of DL for preprocessing MRS data.

Applying Artificial Intelligence to Mitigate Effects of Patient Motion or Other Complicating Factors on Image Quality

Artificial intelligence, particularly deep learning, offers several possibilities to improve the quality or speed of image acquisition in magnetic resonance imaging (MRI). In this article, we briefly review basic machine learning concepts and discuss commonly used neural network architectures for image-to-image translation. Recent examples in the literature describing application of machine learning techniques to clinical MR image acquisition or postprocessing are discussed. Machine learning can contribute to better image quality by improving spatial resolution, reducing image noise, and removing undesired motion or other artifacts. As patients occasionally are unable to tolerate lengthy acquisition times or gadolinium agents, machine learning can potentially assist MRI workflow and patient comfort by facilitating faster acquisitions or reducing exogenous contrast dosage. Although artificial intelligence approaches often have limitations, such as problems with generalizability or explainability, there is potential for these techniques to improve diagnostic utility, throughput, and patient experience in clinical MRI practice.

A modified YOLOv3 detection method for vision-based water surface garbage capture robot

To tackle the water surface pollution problem, a vision-based water surface garbage capture robot has been developed in our lab. In this article, we present a modified you only look once v3-based garbage detection method, allowing real-time and high-precision object detection in dynamic aquatic environments. More specifically, to improve the real-time detection performance, the detection scales of you only look once v3 are simplified from 3 to 2. Besides, to guarantee the accuracy of detection, the anchor boxes of our training data set are reclustered for replacing some of the original you only look once v3 prior anchor boxes that are not appropriate to our data set. By virtue of the proposed detection method, the capture robot has the capability of cleaning floating garbage in the field. Experimental results demonstrate that both detection speed and accuracy of the modified you only look once v3 are better than those of other object detection algorithms. The obtained results provide valuable insight into the high-speed detection and grasping of dynamic objects in complex aquatic environments autonomously and intelligently.

Deep learning-based target metabolite isolation and big data-driven measurement uncertainty estimation in proton magnetic resonance spectroscopy of the brain

PURPOSE

The aim of this study was to develop a method for metabolite quantification with simultaneous measurement uncertainty estimation in deep learning-based proton magnetic resonance spectroscopy (¹ H-MRS).

METHODS

The reliability of metabolite quantification depends on signal-to-noise ratio (SNR), linewidth, and degree of spectral overlap (DSO), and therefore knowledge about these factors may be utilized in measurement uncertainty estimation in deep learning-based ¹ H-MRS. While SNR and linewidth are typically estimated from a representative singlet, DSO needs to be estimated metabolite-specifically. We developed convolutional neural networks (CNNs) capable of isolating target metabolite signal on simulated rat brain spectra at 9.4T, such that, in addition to metabolite content, the signal-to-background ratio (SBR) as a quantitative metric of DSO can be estimated directly from CNN-output for each metabolite. The CNN-predicted SBR was adjusted according to its pre-defined relationship to the ground-truth SBR by exploiting the big spectral data (N = 80 000), and used for measurement uncertainty estimation together with the SNR and linewidth from the CNN-input spectrum. The proposed method was tested first on the simulated spectra in comparison with LCModel and jMRUI and further on in vivo spectra.

RESULTS

The proposed method outperformed LCModel and jMRUI in both quantitative accuracy and measurement uncertainty estimation. Using in vivo data, the metabolite concentrations from the proposed method were close to the reported ranges with the measurement uncertainty of glutamine, glutamate, myo-inositol, N-acetylaspartate, and Tau less than 10%.

CONCLUSION

The proposed method may be used for metabolite quantification with measurement uncertainty estimation in rat brain at 9.4T by exploiting the spectral isolation capability of the CNNs and the availability of big spectral data.

Constrained Magnetic Resonance Spectroscopic Imaging by Learning Nonlinear Low-Dimensional Models

Magnetic resonance spectroscopic imaging (MRSI) is a powerful molecular imaging modality but has very limited speed, resolution, and SNR tradeoffs. Construction of a low-dimensional model to effectively reduce the dimensionality of the imaging problem has recently shown great promise in improving these tradeoffs. This paper presents a new approach to model and reconstruct the spectroscopic signals by learning a nonlinear low-dimensional representation of the general MR spectra. Specifically, we trained a deep neural network to capture the low-dimensional manifold, where the high-dimensional spectroscopic signals reside. A regularization formulation is proposed to effectively integrate the learned model and physics-based data acquisition model for MRSI reconstruction with the capability to incorporate additional spatiospectral constraints. An efficient numerical algorithm was developed to solve the associated optimization problem involving back-propagating the trained network. Simulation and experimental results were obtained to demonstrate the representation power of the learned model and the ability of the proposed formulation in producing SNR-enhancing reconstruction from the practical MRSI data.

An iterative sparse deconvolution method for simultaneous multicolor 19 F-MRI of multiple contrast agents

PURPOSE

19 F-MRI is gaining widespread interest for cell tracking and quantification of immune and inflammatory cells in vivo. Different fluorinated compounds can be discriminated based on their characteristic MR spectra, allowing in vivo imaging of multiple 19 F compounds simultaneously, so-called multicolor 19 F-MRI. We introduce a method for multicolor 19 F-MRI using an iterative sparse deconvolution method to separate different 19 F compounds and remove chemical shift artifacts arising from multiple resonances.

METHODS

The method employs cycling of the readout gradient direction to alternate the spatial orientation of the off-resonance chemical shift artifacts, which are subsequently removed by iterative sparse deconvolution. Noise robustness and separation was investigated by numerical simulations. Mixtures of fluorinated oils (PFCE and PFOB) were measured on a 7T MR scanner to identify the relation between 19 F signal intensity and compound concentration. The method was validated in a mouse model after intramuscular injection of fluorine probes, as well as after intravascular injection.

RESULTS

Numerical simulations show efficient separation of 19 F compounds, even at low signal-to-noise ratio. Reliable chemical shift artifact removal and separation of PFCE and PFOB signals was achieved in phantoms and in vivo. Signal intensities correlated excellently to the relative 19 F compound concentrations (r-2 = 0.966/0.990 for PFOB/PFCE).

CONCLUSIONS

The method requires minimal sequence adaptation and is therefore easily implemented on different MRI systems. Simulations, phantom experiments, and in-vivo measurements in mice showed effective separation and removal of chemical shift artifacts below noise level. We foresee applicability for simultaneous in-vivo imaging of 19 F-containing fluorine probes or for detection of 19 F-labeled cell populations.

Classification and Diagnosis of Thyroid Carcinoma Using Reinforcement Residual Network with Visual Attention Mechanisms in Ultrasound Images

How to differentiate thyroid cancer nodules from a large number of benign nodules is always a challenging subject for clinicians. This paper proposes a novel Sal-deel network model to achieve the classification and diagnosis of thyroid cancer, which can simulate visual attention mechanism. The Sal-deep network introduces saliency map as an additional information on the deep residual network, which selectively enhances the feature extracted from different regions according to the mask map. Sal-deep network can work effectively for the benchmark networks with different data sets and different structures, and it is a universal network model. Sal-deep network increases the complexity of the network, but improves the efficiency of the network. A large number of qualitative and quantitative experiments show that our improved network is superior to other existing deep models in terms of classification accuracy rate and Recall, which is suitable for clinical application.

Applications of Deep Learning to Neuro-Imaging Techniques

Many clinical applications based on deep learning and pertaining to radiology have been proposed and studied in radiology for classification, risk assessment, segmentation tasks, diagnosis, prognosis, and even prediction of therapy responses. There are many other innovative applications of AI in various technical aspects of medical imaging, particularly applied to the acquisition of images, ranging from removing image artifacts, normalizing/harmonizing images, improving image quality, lowering radiation and contrast dose, and shortening the duration of imaging studies. This article will address this topic and will seek to present an overview of deep learning applied to neuroimaging techniques.

Methodological consensus on clinical proton MRS of the brain: Review and recommendations

Proton MRS (1 H MRS) provides noninvasive, quantitative metabolite profiles of tissue and has been shown to aid the clinical management of several brain diseases. Although most modern clinical MR scanners support MRS capabilities, routine use is largely restricted to specialized centers with good access to MR research support. Widespread adoption has been slow for several reasons, and technical challenges toward obtaining reliable good-quality results have been identified as a contributing factor. Considerable progress has been made by the research community to address many of these challenges, and in this paper a consensus is presented on deficiencies in widely available MRS methodology and validated improvements that are currently in routine use at several clinical research institutions. In particular, the localization error for the PRESS localization sequence was found to be unacceptably high at 3 T, and use of the semi-adiabatic localization by adiabatic selective refocusing sequence is a recommended solution. Incorporation of simulated metabolite basis sets into analysis routines is recommended for reliably capturing the full spectral detail available from short TE acquisitions. In addition, the importance of achieving a highly homogenous static magnetic field (B0 ) in the acquisition region is emphasized, and the limitations of current methods and hardware are discussed. Most recommendations require only software improvements, greatly enhancing the capabilities of clinical MRS on existing hardware. Implementation of these recommendations should strengthen current clinical applications and advance progress toward developing and validating new MRS biomarkers for clinical use.

Conditional generative adversarial network for 3D rigid-body motion correction in MRI

PURPOSE

Subject motion in MRI remains an unsolved problem; motion during image acquisition may cause blurring and artifacts that severely degrade image quality. In this work, we approach motion correction as an image-to-image translation problem, which refers to the approach of training a deep neural network to predict an image in 1 domain from an image in another domain. Specifically, the purpose of this work was to develop and train a conditional generative adversarial network to predict artifact-free brain images from motion-corrupted data.

METHODS

An open source MRI data set comprising T₂ *-weighted, FLASH magnitude, and phase brain images for 53 patients was used to generate complex image data for motion simulation. To simulate rigid motion, rotations and translations were applied to the image data based on randomly generated motion profiles. A conditional generative adversarial network, comprising a generator and discriminator networks, was trained using the motion-corrupted and corresponding ground truth (original) images as training pairs.

RESULTS

The images predicted by the conditional generative adversarial network have improved image quality compared to the motion-corrupted images. The mean absolute error between the motion-corrupted and ground-truth images of the test set was 16.4% of the image mean value, whereas the mean absolute error between the conditional generative adversarial network-predicted and ground-truth images was 10.8% The network output also demonstrated improved peak SNR and structural similarity index for all test-set images.

CONCLUSION

The images predicted by the conditional generative adversarial network have quantitatively and qualitatively improved image quality compared to the motion-corrupted images.

Data-driven synthetic MRI FLAIR artifact correction via deep neural network

BACKGROUND

FLAIR (fluid attenuated inversion recovery) imaging via synthetic MRI methods leads to artifacts in the brain, which can cause diagnostic limitations. The main sources of the artifacts are attributed to the partial volume effect and flow, which are difficult to correct by analytical modeling. In this study, a deep learning (DL)-based synthetic FLAIR method was developed, which does not require analytical modeling of the signal.

PURPOSE

To correct artifacts in synthetic FLAIR using a DL method.

STUDY TYPE

Retrospective.

SUBJECTS

A total of 80 subjects with clinical indications (60.6 ± 16.7 years, 38 males, 42 females) were divided into three groups: a training set (56 subjects, 62.1 ± 14.8 years, 25 males, 31 females), a validation set (1 subject, 62 years, male), and the testing set (23 subjects, 57.3 ± 20.4 years, 13 males, 10 females).

FIELD STRENGTH/SEQUENCE

3 T MRI using a multiple-dynamic multiple-echo acquisition (MDME) sequence for synthetic MRI and a conventional FLAIR sequence.

ASSESSMENT

Normalized root mean square (NRMSE) and structural similarity (SSIM) were computed for uncorrected synthetic FLAIR and DL-corrected FLAIR. In addition, three neuroradiologists scored the three FLAIR datasets blindly, evaluating image quality and artifacts for sulci/periventricular and intraventricular/cistern space regions.

STATISTICAL TESTS

Pairwise Student's t-tests and a Wilcoxon test were performed.

RESULTS

For quantitative assessment, NRMSE improved from 4.2% to 2.9% (P < 0.0001) and SSIM improved from 0.85 to 0.93 (P < 0.0001). Additionally, NRMSE values significantly improved from 1.58% to 1.26% (P < 0.001), 3.1% to 1.5% (P < 0.0001), and 2.7% to 1.4% (P < 0.0001) in white matter, gray matter, and cerebral spinal fluid (CSF) regions, respectively, when using DL-corrected FLAIR. For qualitative assessment, DL correction achieved improved overall quality, fewer artifacts in sulci and periventricular regions, and in intraventricular and cistern space regions.

DATA CONCLUSION

The DL approach provides a promising method to correct artifacts in synthetic FLAIR.

LEVEL OF EVIDENCE

4 Technical Efficacy: Stage 1 J. Magn. Reson. Imaging 2019;50:1413-1423.

Intact metabolite spectrum mining by deep learning in proton magnetic resonance spectroscopy of the brain

PURPOSE

To develop a robust method for brain metabolite quantification in proton magnetic resonance spectroscopy (¹ H-MRS) using a convolutional neural network (CNN) that maps in vivo brain spectra that are typically degraded by low SNR, line broadening, and spectral baseline into noise-free, line-narrowed, baseline-removed intact metabolite spectra.

METHODS

A CNN was trained (n = 40 000) and tested (n = 5000) on simulated brain spectra with wide ranges of SNR (6.90-20.74) and linewidth (10-20 Hz). The CNN was further tested on in vivo spectra (n = 40) from five healthy volunteers with substantially different SNR, and the results were compared with those from the LCModel analysis. A Student t test was performed for the comparison.

RESULTS

Using the proposed method the mean-absolute-percent-errors (MAPEs) in the estimated metabolite concentrations were 12.49% ± 4.35% for aspartate, creatine (Cr), γ-aminobutyric acid (GABA), glucose, glutamine, glutamate, glutathione (GSH), myo-Inositol (mI), N-acetylaspartate, phosphocreatine (PCr), phosphorylethanolamine, and taurine over the whole simulated spectra in the test set. The metabolite concentrations estimated from in vivo spectra were close to the reported ranges for the proposed method and the LCModel analysis except mI, GSH, and especially Cr/PCr for the LCModel analysis, and phosphorylcholine to glycerophosphorylcholine ratio (PC/GPC) for both methods. The metabolite concentrations estimated across the in vivo spectra with different SNR were less variable with the proposed method (~10% or less) than with the LCModel analysis.

CONCLUSION

The robust performance of the proposed method against low SNR may allow a subminute ¹ H-MRS of human brain, which is an important technical development for clinical studies.