Search

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.

This literature review presents a comprehensive overview of machine learning (ML) applications in proton magnetic resonance 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 magnetic resonance 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.

These imaging data are examples of sub-millimeter resolution T1-weighted EPI (Echo Planar Imaging) acquired using the T123DEPI (T1-imaging with 2 3D-EPIs) sequence [1]; functional MRI data with matching resolution and distortion, and MP2RAGE (Magnetization Prepared 2 Rapid Acquisition Gradient Echoes) anatomical images [2], from the same subjects. Data from two protocols and subjects presented in the paper describing the sequence [1] are made available here: 1)0.8 mm protocol: whole brain, axial T123DEPI T1-weighted images; a 5-minute fMRI run with the same orientation and 27 mm coverage in the slice selection direction, covering the primary visual cortex. fMRI data were acquired while the volunteer viewed a flashing checkerboard stimulus; the unsmoothed GLM results of the fMRI and a 0.64 mm resolution MP2RAGE from the same subject. These data are from Experiment 3 in [1] 2)0.7 mm protocol: partial brain T123DEPI T1-weighted images with longer or shorter readouts; matching coronal echo planar images again acquired while viewing a flashing checkerboard stimulus and a 0.64 mm whole brain MP2RAGE from the same subject. These data are from Experiment 1 in [1].

The increasing availability of ultra-high field scanners has led to a growing number of submillimetre fMRI studies in humans, typically targeting the gray matter at different cortical depths. In most analyses, the definition of surfaces at different cortical depths is based on an anatomical image with different contrast and distortions than the functional images. Here, we introduce a novel sequence providing bias-field corrected T1-weighted images and T1-maps with distortions that match those of the fMRI data, with an image acquisition time significantly shorter than standard T1-weighted anatomical imaging. For 'T1-imaging with 2 3D-EPIs', or T123DEPI, 3D-EPI volumes are acquired centred at two inversion times. These 3D-EPIs are segmented into half, quarter or smaller blocks of k-space to allow for optimisation of the inversion times. T1-weighted images and T1-maps are then generated as for MP2RAGE acquisitions. A range of T123DEPI data acquired at 7 T is shown with resolutions ranging from 0.7 mm to 1.3 mm isotropic voxels. Co-registration quality to the mean EPI of matching fMRI timecourses shows markedly less local deviations compared to co-registration of a standard MP2RAGE to the same echo planar volume. Thus, the T123DEPI T1-weighted images and T1-maps can be used to provide cortical surfaces with matched distortions to the functional data or else to facilitate co-registration between functional and undistorted anatomical data.

These imaging data are examples of sub-millimeter resolution T1-weighted EPI (Echo Planar Imaging) acquired using the T123DEPI (T1-imaging with 2 3D-EPIs) sequence [1]; functional MRI data with matching resolution and distortion, and MP2RAGE (Magnetization Prepared 2 Rapid Acquisition Gradient Echoes) anatomical images [2], from the same subjects. Data from two protocols and subjects presented in the paper describing the sequence [1] are made available here: 1)0.8 mm protocol: whole brain, axial T123DEPI T1-weighted images; a 5-minute fMRI run with the same orientation and 27 mm coverage in the slice selection direction, covering the primary visual cortex. fMRI data were acquired while the volunteer viewed a flashing checkerboard stimulus; the unsmoothed GLM results of the fMRI and a 0.64 mm resolution MP2RAGE from the same subject. These data are from Experiment 3 in [1] 2)0.7 mm protocol: partial brain T123DEPI T1-weighted images with longer or shorter readouts; matching coronal echo planar images again acquired while viewing a flashing checkerboard stimulus and a 0.64 mm whole brain MP2RAGE from the same subject. These data are from Experiment 1 in [1].

The increasing availability of ultra-high field scanners has led to a growing number of submillimetre fMRI studies in humans, typically targeting the gray matter at different cortical depths. In most analyses, the definition of surfaces at different cortical depths is based on an anatomical image with different contrast and distortions than the functional images. Here, we introduce a novel sequence providing bias-field corrected T1-weighted images and T1-maps with distortions that match those of the fMRI data, with an image acquisition time significantly shorter than standard T1-weighted anatomical imaging. For 'T1-imaging with 2 3D-EPIs', or T123DEPI, 3D-EPI volumes are acquired centred at two inversion times. These 3D-EPIs are segmented into half, quarter or smaller blocks of k-space to allow for optimisation of the inversion times. T1-weighted images and T1-maps are then generated as for MP2RAGE acquisitions. A range of T123DEPI data acquired at 7 T is shown with resolutions ranging from 0.7 mm to 1.3 mm isotropic voxels. Co-registration quality to the mean EPI of matching fMRI timecourses shows markedly less local deviations compared to co-registration of a standard MP2RAGE to the same echo planar volume. Thus, the T123DEPI T1-weighted images and T1-maps can be used to provide cortical surfaces with matched distortions to the functional data or else to facilitate co-registration between functional and undistorted anatomical data.

Purpose To compare methods for estimating B0 maps used in retrospective correction of high-resolution anatomical images at ultra-high field strength. The B0 maps were obtained using three methods: (1) 1D navigators and coil sensitivities, (2) field probe (FP) data and a low-order spherical harmonics model, and (3) FP data and a training-based model. Methods Data from nine subjects were acquired while they performed activities inducing B0 field fluctuations. Estimated B0 fields were compared with reference data, and the reductions of artifacts were compared in corrected T2* images. Results Reduction of sum-of-squares difference relative to a reference image was evaluated, and Method 1 yielded the largest artifact reduction: 27 ± 15%, 20 ± 18% (mean ± 1 standard deviation) for deep breathing and combined deep breathing and hand motion activities. Method 3 performed almost as well (24 ± 18%, 15 ± 17%), provided that adequate training data were used, and Method 2 gave a similar result (21 ± 16%, 19 ± 17%). Conclusion This study confirms that all of the investigated methods can be used in retrospective image correction. In terms of image quality, Method 1 had a small advantage, whereas the FP-based methods measured the B0 field slightly more accurately. The specific strengths and weaknesses of FPs and navigators should therefore be considered when determining which B0-estimation method to use.

Results: We included 6 patients. The age range of the patients was 35–65 years and 5 of them were women. 1 out of 6 had a 1.5 T MRI, the other 5 patients had a 3 T MRI previous to the 7 T MRI. The lesion size varied between 0.9 mm and 2.0 mm. In 5 of the 6 patients the presence of an infundibulum could be proven using the high resolution of the 7 T MRA. All patients tolerated the 7 T MRI well. Conclusion: Our results suggest that high resolution and contrast of 7 T MRA provides added diagnostic value in discriminating between intracranial aneurysms and infundibula. This finding may have important consequences for patient follow-up and comfort because it might reduce unnecessary follow-up exams and decrease uncertainty about the diagnosis. Larger studies, however, are needed to confirm our findings.