Keywords: AI/ML Image Reconstruction, Brain

Motivation: In MRI reconstruction, deep-learning methods often increase network complexity for improved super-resolution, leading to longer reconstruction times and training difficulties.

Goal(s): Our solution introduces an enhanced lightweight network that maintains high-quality performance. 

Approach: We accomplish this by stacking Reverse Residual Attention Fusion (RRAF) with PCA and Enhanced Spatial Attention (ESA) for precise feature extraction, utilizing Transformers with depth-wise dilated convolution for better context information, and employing High-Frequency Image Refinement (HFIR) for detailed information recovery.

Results: Our experiments confirm the effectiveness of our approach.

Impact: Introducing the lightweight network represents an important improvement in MRI SR reconstruction. By integrating Reverse Residual Attention Fusion, it upholds exceptional image quality, streamlines network complexity, reduces reconstruction time, and simplifies training for SR MRI image reconstruction.

Keywords: Analysis/Processing, Relaxometry

Motivation: Quantitative MRI has the potential for improved disease characterization, but the limited accessibility impedes its application.

Goal(s): To develop a deep learning method for retrospective T1 and T2 quantification from real-world brain MRI data, with the ability to handle diverse imaging protocols.

Approach: A protocol-aware self-supervised learning framework was developed, with the imaging parameters incorporated as additional inputs to the model.

Results: Validation on volunteers showed errors within 10% for nine brain regions when compared to prospective T1/T2 mapping. Application to 376 glioblastoma patients with diverse imaging protocols revealed statistical differences in T1 and T2 among tumor sub-regions and normal-appearing tissues.

Impact: The proposed method may allow retrospective T1 and T2 mapping in large real-world MRI datasets, enabling analysis of them regardless of the difference in protocols and scanners. This will facilitate the large-scale investigation of quantitative MRI as biomarkers for diseases.

Keywords: Analysis/Processing, Machine Learning/Artificial Intelligence

Motivation: Traditional medical image reconstruction emphasizes standard metrics, potentially overlooking optimization for downstream tasks like segmentation and anomaly detection.

Goal(s): Our study investigates the relationship between standard reconstruction and object detection metrics.

Approach: We trained a Faster R-CNN detector for meniscal anomalies, addressing class imbalance and implementing a custom detection-specific augmentation protocol.

Results: Evaluation on reconstructed datasets revealed that reconstruction quality was associated with true predictions but had a limited impact on overall detection performance, while boxes-based reconstruction metrics showed no correlation with prediction outcomes. These findings underscore the importance of considering associations between standard reconstruction and downstream task metrics when optimizing end-to-end pipelines.

Impact: Evaluation of standard reconstruction metrics, sliced by object detection outcomes, revealed a significant association between reconstruction and detection performance, emphasizing the utility of this approach in assessing task-based reconstruction.

Keywords: Analysis/Processing, PET/MR, Bowsher CNN

Motivation: Bowsher CNN, an MRI anatomic-guided PET reconstruction method, can provide high-resolution PET images. It is unknown how it may affect lesion characterization in MS patients.

Goal(s): Evaluate PET signal without or with Bowsher CNN in various MS lesions.

Approach: FLAIR images defined WMH lesions and NAWM. WMH lesions were further separated into “T1-hypo” and “T1-iso” sub-categories based on T1 intensity. 18F-FDG SUVs were obtained from various lesion ROIs.

Results: The SUV differences between WMH and NAWM and between T1-hypo and T1-iso lesions became lower and higher, respectively, after applying Bowsher CNN.

Impact: Bowsher CNN PET reconstruction results in high-resolution PET images. SUV differences between WMH and NAWM became smaller, while the SUV differences between T1-hypo and T1-iso, two WMH subcategories, were larger after applying Bowsher CNN. 

Keywords: Analysis/Processing, Data Processing, synthetic CT

Motivation: Standard intensity-based voxel-wise losses, generally used in image-to-image translation techniques, are typically biased towards the estimation of the low frequency content in image spectra. For the generation of synthetic CT (sCT) contrast, this results in limited image sharpness, and consequently a limited clinical utility.

Goal(s): To improve sharpness in synthetic contrasts. 

Approach: We trained a model using a combination of intensity- and frequency-based losses for the generation of sCT images from MRI. 

Results: Compared to a baseline model, sCT images generated using the focal-frequency loss resulted in an enhanced level of details in knee images.

Impact: Our results suggest that the use of frequency-based losses, in conjunction with an intensity-based L1 loss, improves image sharpness in synthetic contrasts, and thereby shows the potential to increase their clinical usefulness.

Keywords: Analysis/Processing, Spinal Cord, Contrast Neutralization

Motivation: Provide flexibility to clinicians to fine-tune protocols/contrasts while still leveraging existing Deep-learning (DL) applications trained with limited set of MR contrasts.

Goal(s): Develop task-specific contrast neutralization pre-processing step to handle multiple imaging contrasts, that are different from the contrasts in the trainset.

Approach: Investigate Simple Contrast Neutralization (SCNe) approach that leverages Fourier domain filtering to neutralize contrast from objects of desired sizes, and demonstrate its impact on generalization of cervical foramina plane determination.

Results: Statistically significant improvements in prediction of planes when SCNe is used on new MERGE T2* contrast with DL-model that was trained only with Ax-T2 images.

Impact: Use of our Simple Contrast Neutralization (SCNe) approach as pre-processing step was effective in making DL-model trained only with Ax-T2 images robust to unseen new contrast MERGE T2* MR dataset for Spine cervical foramina (CF) plane determination.

Keywords: Analysis/Processing, Machine Learning/Artificial Intelligence, multi-frequency Gibbs artifact, deep learning, model training, artifact removal

Motivation: Gibbs artifact generated by zero-padding k-space data for model training poses a huge challenge for the model to learn different severity and manifestation of Gibbs artifact in the image domain.

Goal(s): Our goal was to effectively remove ringing artifact with a deep-learning model by developing a novel multi-frequency Gibbs generator algorithm.

Approach: We introduced Gibbs artifact generator (GAG) algorithm to create Gibbs artifacts with different truncation ratios as the input and tested the performance with a proposed deep-learning model. 

Results: The images processed using the proposed approach demonstrated higher image quality score than the original images (all P < 0.05).

Impact: The images generated by our new GAG algorithm with pronounced multi-frequency Gibbs artifacts could be used as a reliable training set for deep-learning model training, enabling the model to effectively identify and eliminate Gibbs artifacts in spinal MR imaging.

Keywords: Analysis/Processing, Low-Field MRI, ULF MRI, Ultra-Low-Field MRI, Deep Learning, Unpaired Image Translation, Brain Segmentation, Vision Transformers, CycleGAN

Motivation: The image quality of ultra-low-field MRI impacts the reliability of volumetric analysis in the brain. Existing techniques that address this issue learn from synthetically generated images, leading to a domain shift problem when presented with real images.

Goal(s): Development of a deep learning method trained with real ULF and HF images to robustly generate an image that can be segmented with routine software tools.

Approach: We introduce a CycleGAN framework with Residual Vision Transformers to improve super-resolved images compared to existing methods.

Results: The accuracy of volumetric estimations improves using our method compared to others based on clinical correlations and test-retest reliability metrics.

Impact: Our new image enhancement method should allow reliable volumetric evaluation using ULF-MRI. This will allow investigators in regions with access to ULF systems to monitor brain health in a way that was previously unattainable.

Keywords: Analysis/Processing, Cancer, Cervical Cancer，Dynamic Contrast-Enhanced MRI，UNet，Vision Transformer, CycleGAN, self-supervised pretraining

Motivation: DCE-MRI plays an important role in non-invasive detection and monitoring of cervical cancer, providing key information for improving diagnosis and treatment accuracy.

Goal(s): DCE-MRI faces complexities and noise issues in application and needs to be optimized and improved by deep learning techniques for parameter mapping. Existing deep learning based methods suffer from limited data and model efficiency.

Approach: We propose a CycleGAN-like model with UNet-Vision-Transformer generator, enhance the discriminator with gradient penalty, and pre-train the model via self-supervised image inpainting.

Results: The numerical experimental results demonstrate that the proposed model is quite efficient and robust compared with other deep learning-based methods.

Impact: This research offers fresh avenues for processing medical imaging data by proposing a novel and efficient deep learning model, significantly impacting the improvement of disease diagnosis. Furthermore, it provides researchers with new directions and insights, advancing scientific and technological progress.

Keywords: Analysis/Processing, Neuro

Motivation: For subcortical brain segmentation, the most widely accepted tools like FreeSurfer are slow and inefficient for large datasets, while faster methods often sacrifice accuracy and reliability.

Goal(s): In this study, we propose a novel deep learning based alternative and achieve consistent state-of-the-art performance within reasonable processing times.

Approach: Our model, TABSurfer, utilizes a 3D patch-based approach with a hybrid CNN-Transformer architecture.

Results: We evaluated TABSurfer against FreeSurfer ground truths across various T1w MRI datasets, consistently demonstrating strong performance over a leading deep learning benchmark, FastSurferVINN. Then, we validated TABSurfer on a manual reference, outperforming both FreeSurfer and FastSurferVINN based on the gold standard.

Impact: Our proposed deep learning model, TABSurfer, demonstrated state-of-the-art subcortical segmentation performance and utility. TABSurfer displayed reliability across numerous datasets and outperformed well established traditional and deep learning tools in FreeSurfer and FastSurferVINN.

Keywords: Analysis/Processing, fMRI (resting state), Functional Connectivity, Graph Kernel, Brain connectivity, Signal Modeling, Signal Representations

Motivation: Various rs-fMRI studies highlight the need for accurate delineation of different brain functional networks (FNs) to carry out precise therapeutic interventions in the individuals.

Goal(s): To develop a novel zero-shot non-linear graph kernel-assisted approach for enhanced functional brain parcellation at individual and group levels.

Approach: Utilization of Wavelet, Fourier, and Hilbert transformations for feature extraction from BOLD signals, and a propagation attribute graph kernel to capture non-linear temporo-spatial connectivity, using k-means clustering.

Results: The kernel-based approach outperforms static FC matrix parcellations, achieving higher accuracy in network delineation in both individual and group level, as evidenced by Dice and Jaccard scores.

Impact: The study introduced graph kernel-based method for functional brain parcellation, which improved the accuracy of functional network delineation in rs-fMRI data, surpassing traditional static functional connectivity approaches in both individual and group level, as validated by Dice and Jaccard metrics.

Keywords: Analysis/Processing, Segmentation, Left Atrium, Mitral Valve Regurgitation, Domain Adaptation, Deep Learning

Motivation: Addressing challenges with current deep learning (DL) techniques that struggle with domain shifts.

Goal(s): To introduce a domain-adaptive technique that is able to segment the Left Atrium from MRI of patients employing model trained exclusively on healthy data.

Approach: Our approach involves training exclusively on healthy data and incorporating stochastic encoding of temporal composite variations as augmentations to encode the underlying space of plausible anatomical changes and dynamics. We tested on three challenging unseen patient daatsets.

Results: Our domain-adaptive approach showed significant improvement over the state-of-the-art LA segmentation model. Enabling LA segmentation of all time frames of the cardiac cycle.

Impact: The proposed domain-adaptive deep learning approach addresses a fundamental challenge of training deep learning models only on healthy control datasets while maintaining high performance on unseen patients' populations. This could potentially lead to solve performance issues for limited patients cohorts.  

Keywords: Analysis/Processing, Reproductive

Motivation: Conventional deep learning-based harmonization cannot handle the unseen source domain image when there is no large-size data.

Goal(s): We propose blind harmonization, which requires only target domain data during training and generalizes well on unseen source domain data.

Approach: BlindHarmony utilizes an unconditional flow model to measure the probability of the target domain image and find a harmonized image that is structurally close to this source domain image but has a high probability in the target domain.

Results: BlindHarmony successfully harmonized the source domain image to the target domain and improved the performance of downstream tasks for the data with a domain gap.

Impact: Deep learning-based harmonization typically necessitates both source and target domain data, limiting its widespread applicability. This study eliminates the need for source domain data and exhibits robust generalization to new source domain data, thereby expanding the utility of harmonization.

Keywords: Analysis/Processing, Quantitative Imaging

Motivation: Tissue properties are estimated from MRI data using bio-physical models that relate MRI signal to underlying tissue properties via quantitative MRI parameters. Deep learning can improve parameter estimation, but needs retraining for different acquisition protocols, hindering implementation. 

Goal(s): Implement a deep learning algorithm able to estimate quantitative MRI parameters for multiple quantitative MRI applications, irrespective of acquisition protocol.  

Approach: Neural controlled differential equations (NCDEs) overcome this limitation as they are independent of the configuration of input data.

Results: NCDEs have improved performance compared to least squares minimization in estimating quantitative MRI parameters when SNR is low or when the parameter has low sensitivity. 

Impact: Neural controlled differential equations are a generic purpose tool for parameter estimation in quantitative MRI that outperform least squares minimization in quantitative MRI parameter estimation, irrespective of acquisition protocol or quantitative MRI application. 

Keywords: Analysis/Processing, Quantitative Susceptibility mapping

Motivation: Deep learning approaches for QSM-based dipole inversion lack generalizability towards acquisition parameters.

Goal(s): Our aim was to address data scarcity by integrating known information in the network model and investigate the feasibility of transfer learning.

Approach: The acquisition parameters (voxel size, FOV orientation) were integrated with manifold learning. The models were pre-trained on large-scale synthetic data sets and fine-tuned on in-vivo brain data in a second step.

Results: The use of manifold learning increased generalizability, while transfer learning substantially improved the quality of computed susceptibility maps.

Impact: While this study demonstrates the feasibility of cross-domain knowledge transfer in deep learning approaches for QSM, it also points to the potential of fine-tuning network parameters to scanner-specific data in general, boosting the performance of neural networks therewith.

Keywords: Analysis/Processing, Segmentation

Motivation: Accurate segmentation of free-breathing (FB) myocardial perfusion (MP) MRI is a labor-intensive yet necessary preprocessing step. A quality control (QC) tool for deep learning (DL)-based segmentation of FB MP MRI is lacking.

Goal(s): Developing a DL-based dynamic QC (dQC) tool for automatic analysis of MP MRI.

Approach: Using the discrepancy between patch-based segmentations, a dQC map is derived and quantified into a dQC metric. The utility of this metric in detecting erroneous segmentations is demonstrated by considering a human-in-the-loop (HiTL) framework.

Results: Referral of the dQC-detected timeframes to a HiTL has markedly improved the segmentation results when compared to a random referral approach.

Impact: We proposed a dynamic quality control tool for automatic segmentation and analysis of free-breathing myocardial perfusion MRI datasets. Our results show that the proposed approach has markedly improved segmentation accuracy when used within a practical and efficient clinician-in-the-loop setting.

Keywords: Analysis/Processing, Spectroscopy, ML Robustness, MRS Quantification

Motivation: Despite promising developments, current machine learning methods for magnetic resonance spectroscopy (MRS) suffer from limited robustness and generalization issues, restricting their clinical application.

Goal(s): This study compares training strategies for MRS quantification, focusing on neural network resilience to out-of-distribution samples. 

Approach: Bias towards the training distribution was assessed for various out-of-distribution cases in synthetic data and in-vivo data. 

Results: Our findings reveal that, while common supervised regression is most accurate for in-distribution cases, it shows the most data bias; physics-informed self-supervised training is more robust; while integrating a least-squares fitting method within the training framework enhances standalone performance while remaining generalizable.

Impact: To advance integration in clinical MRS, robust and generalizable machine learning methods are needed. This study's exploration of quantification training strategies offers insights into data biases and advocates hybrid models that combine traditional methods with neural networks to maintain robustness.

Keywords: Analysis/Processing, Hyperpolarized MR (Gas), Deep Learning; Magnetic Resonance Imaging (MRI); Hyperpolarized Gas MRI; Segmentation; Ventilation Defect; Chronic Obstructive Pulmonary Disease (COPD); Lung Imaging

Motivation: Current methods for quantifying lung ventilation defects using hyperpolarized gas MRI are effective but time-consuming. Deep Learning offers potential enhancements in image segmentation, with Vision Transformers (ViTs) emerging as notable alternatives to traditional CNNs.

Goal(s): The study aims to assess SegFormer's capability for automating the segmentation and quantification of ventilation defects in hyperpolarized gas MRI, comparing its efficiency and accuracy against traditional methods.

Approach: Utilizing a dataset from 56 study participants, the study adopted the SegFormer architecture for segmenting MRI slices.

Results: SegFormer, especially with ImageNet pretraining, surpassed CNN-based techniques in segmentation. Specifically, the MiT-B2 configuration of SegFormer showcased exceptional efficacy and efficiency.

Impact: SegFormer's efficiency in hyperpolarized gas MRI enhances future clinical decision-making with swift and precise segmentation. Its superiority may inspire broader adoption and further exploration into Vision Transformers' potential in medical imaging.

Keywords: Analysis/Processing, Simulations, Motion artifacts

Motivation: Motion compromises the utility of structural MRI with MP-RAGE sequence, a workhorse of quantitative neuroimaging research. Recent interest in deep learning-based mitigating solutions, and the scarcity of motion-corrupted data, motivates the need for realistic data simulation. Unfortunately, existing open-source simulators fail to consider important features in real-world acquisitions, including variations in phase-encoding direction, multi-coil acquisition and GRAPPA parallel imaging, resulting in less realistic simulations.

Goal(s): We aim to develop a more realistic motion artifact simulator for MP-RAGE structural MRI.

Approach: We extend TorchIO, an existing simulation framework, to support aforementioned features.

Results: The comparison between simulations demonstrated the importance of including these features.

Impact: The proposed simulation framework can be used to generate more realistic motion-corrupted MRI data from clean images. These data can be served as training sets for deep learning algorithms in motion artifact related applications.