Keywords: Analysis/Processing, Machine Learning/Artificial Intelligence, Image Registration, Multimodal, Minimal-Invasive, Image-Guided Intervention

Motivation: We developed neural networks for deformable medical image registration using multiple steps and resolutions.

Goal(s): To investigate how multiresolution networks impact registration results compared to monostep-monoresolution networks.

Approach: The networks were trained unsupervised with Mutual Information and Gradient L2 loss. We compared them with a monoresolution-monostep network and the classical registration method SimpleElastix. We evaluated the multistep networks using a three-dimensional liver dataset with CT and T1-weighted MR scans.

Results: Incorporating multiple steps and resolutions in the neural network yielded registration results with high spatial alignment and medically plausible transformations (minimal image folding) and fast registration times of less than half a second.

Impact: Since the inclusion of multiple steps and resolutions within the neural network leads to improved registration results, multistep registration methods should be used whenever possible. Consequently, more work should be invested in developing multistep-multiresolution networks for multimodal medical image registration.

Keywords: Analysis/Processing, PET/MR, Multi-model Registration, Dual Attention Mechanism

Motivation: Registered PET-MRI is better than single modality in diagnoses, and traditional algorithms are time-consuming and perform poorly in cross-modal registration.

Goal(s): Improve registration efficiency and reduce registration time by improving traditional deep learning networks.

Approach: We propose a weakly-supervised PET-MRI registration network based on a hybrid adaptive attention mechanism. Masks extracted from the fine-tuned large model is uesd to constrain the network.

Results:  We validate the proposed method on liver PET-MRI images. The experimental results show that the proposed method achieves a higher DICE value and shorter registration time than the other state-of-the-art registration algorithms.

Impact: Our proposed new network can help doctors to complete the registration between PET and MRI and diagnose a disease in a short period of time.

Keywords: Analysis/Processing, Motion Correction, MOCO, VoxelMorph

Motivation: Image registration followed by averaging is a common technique to improve the quality of free-breathing single-shot cardiac images. However, registering images becomes challenging when the SNR is low.

Goal(s): Improve image registration for free-breathing cardiac MRI.

Approach: We train a network, called AvgMorph, to register all source images to one target image. In addition, we use the output of a sophisticated deep learning-based edge detector to compute loss.

Results: We validate AvgMorph using a realistic MRXCAT digital phantom for late gadolinium enhancement. AvgMorph outperforms existing methods in terms of NMSE, SSIM, and perceptual quality metrics.

Impact: Pairwise registration of free-breathing images is suboptimal. We propose a network to register all source images to a single target image and utilize a loss function computed to edge maps rather than the images themselves.

Keywords: AI/ML Image Reconstruction, Data Analysis

Motivation: Many deep learning (DL) reconstruction models do not assess the reconstructed phase despite its importance in certain imaging techniques.

Goal(s): To develop a phase-specific metric and demonstrate its suitability for reconstruction assessment.

Approach: We used our developed metric to assess and analyze DL and non-DL reconstruction results in an experiment investigating the effect of coil overlap on DL reconstruction methods. The phase metric results were compared to magnitude metric results.

Results: The phase metric results were consistent with the magnitude metric results and provided useful insights into model performance.

Impact: We propose and test a phase-specific metric that can be used to assess and further the development of complex-valued DL reconstruction methods. This metric would allow for DL reconstruction methods to be applied to MR imaging techniques such as phase contrast imaging.

Keywords: AI/ML Image Reconstruction, Vessels

Motivation: Addressing the challenge of segmenting cerebral vessels in TOF-MRA images, we explored transfer learning to overcome the need for large, annotated datasets.

Goal(s): Assess the feasibility of using a refined CTA-based 3D CNN model for MRA vascular segmentation with a limited dataset.

Approach: Implemented transfer learning on a ResU-Net3 model, initially trained on CTA scans, fine-tuned with a small MRA dataset.

Results: Post-transfer learning, the model's DSC improved dramatically, indicating effective MRA vessel segmentation with limited data.

Impact: This study benefits radiologists by streamlining the segmentation of cerebral vessels in MRA, reducing the workload associated with annotation. The method has the potential to be integrated into clinical workflows, enhancing the efficacy of vascular reconstruction in clinical settings.

Keywords: AI/ML Image Reconstruction, Image Reconstruction, 2D T2 FLEX, Spine

Motivation: Deep learning reconstruction in spine imaging has become widely available, but diagnostic quality and quantitative measurements still need to be verified in the clinical setting. 

Goal(s): Our goal was to validate application of deep learning to 2D FLEX spine imaging at 1.5T via IQ assessment and noise characterization.

Approach: DL and conventionally reconstructed images were assessed by three radiologists. Noise characteristics were evaluated by calculation of total variation and number of detected edges. Fat fraction was also calculated.

Results: The radiologists preferred DL in majority of cases (79.5%), with noise noticeably lower in DL images. The fat fraction measurements were very similar. 

Impact: The application of DL to routine 2D FLEX imaging in the spine provides enhanced diagnostic quality and decreased noise without sacrificing quantitative fidelity. Therefore, the application of DL can be confidently used at 1.5T in the clinic for patient care.

Keywords: AI/ML Image Reconstruction, Machine Learning/Artificial Intelligence, Sodium MRI, MRI, Convolutional Neural Networks, Image Reconstruction

Motivation: Sodium Magnetic Resonance Imaging (²³Na MRI) provides unique metabolic information but suffers from low signal-to-noise ratio (SNR). Iterative anatomically guided reconstructions (AGR) can improve SNR and resolution but are limited in practice by their long computational times.  

Goal(s): To address these limitations, we explore the use of neural networks to approximate the AGR sodium MRI reconstruction and reduce computational time. 

Approach: A U-Net convolutional neural network (CNN) was trained to approximate the AGR iterative reconstruction using data from normal human volunteers. 

Results: Our results indicate that the neural network implementation achieves comparable image quality while significantly reducing reconstruction time.

Impact: The improved SNR accuracy and spatial resolution of the CNN AGR reconstructions make the use of Sodium MRI more feasible within the confines of a clinical examination.  

Keywords: AI/ML Image Reconstruction, Machine Learning/Artificial Intelligence

Motivation: MRI examination of temporomandibular joint takes a long time to scan, especially when scanning in the open-mouth position, the patient cannot maintain the open-mouth state for a long time, it is easy to cause movements of the temporomandibular joint, resulting in motion artifacts in the scanned image and failure of the examination. On the premise of ensuring image quality, it is particularly important to reduce scan time.

Goal(s): Applying deep learning reconstruction to improve temporomandibular joint MRI image quality and reduce scanning time.

Approach: clinical experiments.

Results: Deep learning reconstruction can improve image quality and significantly shorten scanning time of temporomandibular joint MRI.

Impact: Deep learning reconstruction can significantly shorten scan time while improving image quality, helping radiologists to diagnose quickly and confidently，helping patients to spend less time feeling anxious in an MRI and more time living their life.

Keywords: AI/ML Image Reconstruction, Machine Learning/Artificial Intelligence

Motivation: Super-resolution is challenging especially in medical imaging where recovering fine structures from low-resolution images is essential for accurate diagnosis.

Goal(s): Our goal is to achieve super-resolution for the low-resolution knee MR images.

Approach: We propose an edge map reinforced super-resolution GAN with attention-based DenseNet generator to enhance the knee structures in two stages: (1) Generating an enhanced edge map for the low-resolution input image; (2) Generating super-resolution output by using the enhanced edge map generated in the first stage and low-resolution input images.

Results: The results show that the proposed method provides better output than the contemporary super-resolution GAN method.

Impact: This work presents a robust edge map reinforced super-resolution framework for knee MRI by using an attention-based DenseNet generator in super-resolution GAN. The proposed method generates MR images with clear edges and facilitates accurate medical assessment for knee related disorders.

Keywords: AI/ML Image Reconstruction, Normal development, Pituitary

Motivation: 3D CUBE imaging of pediatric pituitary is time-consuming, and thus presents difficulties for children with limited patience and cooperation. A vendor-provided deep learning reconstruction (DLR) algorithm, proposed for high image SNR, may allow for MR imaging with shortened scan time.

Goal(s): Explore if DLR allowed for rapid CUBE imaging in pediatric pituitary while maintaining the image quality and precise measurement of pituitary height.

Approach:  The imaging quality, scan time, and pituitary height measured were compared between DLR-CUBE and conventional CUBE.

Results: Relative to conventional CUBE, DLR-CUBE showed improved SNR, comparable image quality, accurate measurement of pituitary height, and only half the scan time.

Impact: DLR-CUBE can dramatically shorten the acquisition time while maintaining the image quality and accurate measurement for pituitary height, demonstrating the potential of DLR-CUBE in clinical examinations of pediatric pituitary.

Keywords: AI/ML Image Reconstruction, Validation

Motivation: SmartSpeed AI, deep learning-based MR image reconstruction method can be used for scan acceleration, but its clinical applicability for studying brain volumetry and/or cortical myelin has not been investigated. 

Goal(s): This study was aimed to quantitatively evaluate the reliability for estimates of cortical thickness and myelin estimated from SmartSpeed AI reconstruction.

Approach: Segmentation performance was evaluated using Dice coefficient and Hausdorff distance and the reliability of estimation for cortical thickness and myelin was assessed using intraclass correlation coefficient. 

Results: Comparable segmentation accuracy and reliable estimates of cortical thickness and myelin were obtained from relatively high acceleration factor with SmartSpeed AI reconstruction. 

Impact: SmartSpeed AI reconstruction enabled accurate cortical segmentation, and the reliable estimation of cortical thickness and intracortical myelin, suggesting the validity of its clinical applicability with reduced scan time.

Keywords: Data Processing, Machine Learning/Artificial Intelligence

Motivation: Existing phase processing methods  often require users to trade off between time and precision. Therefore, a phase processing network that can dynamically activate different parts based on input samples is of great research value.

Goal(s): We hope that the proposed network can adaptively determine whether to begin with VOI extraction, i.e., removing the brain skull, and provide different solutions for samples of different complexity. 

Approach: We combine dynamic neural network and deformable convolution in the network design to realize dynamic activation and verify it on MRI phase data.

Results: Our dynamic activation based network (DANet) implements adaptive phase processing and achieve competitive performance.

Impact: Our methodological framework can be applied across various field related to phase signal processing, such as Optical Interferometry (OI), Magnetic Resonance Imaging (MRI), Fringe Projection Profilometry (FPP), and Interferometric Synthetic Aperture Radar (InSAR).

Keywords: AI/ML Image Reconstruction, Image Reconstruction, Image Harmonization

Motivation: In multi-centric, multi-scanner datasets differences between images caused by different acquisition protocols and reconstruction techniques are found. These differences introduce biases when developing AI-based solutions that limit their generalization.

Goal(s): To develop an algorithm to harmonize intensities on MR medical images independently of the source and artifacts that may be introducing a bias.

Approach: Use of the MRI frequency domain to synthetically generate realistic intensity variations simulating differences in acquisition protocols. Image-to-image CNN-based solution to reconstruct any image to a reference dataset.

Results: Harmonization of prostate T2w MRI showed a qualitative harmonization of the images and an improvement in AI-based segmentation task.

Impact: This methodology helps the harmonization of medical MRI images, enhancing accuracy and efficiency in MRI AI based task. By standardizing image quality and reducing variations, this innovation ensures consistent interpretations across healthcare institutions, improving collaboration among medical and AI professionals.

Keywords: AI/ML Software, Segmentation

Motivation: PREFUL MRI relies on accurate lung segmentation. Automating lung segmentation using supervised machine learning requires the laborious creation of training data. Therefore, an alternative independent of availability and peculiarities of training data may be useful.  

Goal(s): Investigate feasibility and limits of lung segmentation in PREFUL MRI across different vendors, acquisition parameters, age groups and pulmonary diseases without training data. 

Approach: Segment Anything Model (SAM) using different point grids and seedpoint-based prompts was evaluated in overall 14 different configurations. 

Results: Comparison with ground truth showed median Dice Similarity Coefficient (DSC) of 0.82 without training data.  

Impact: Lung segmentation of PREFUL MRI of child and adult patients with different pulmonary diseases appears feasible without training data. The construction of supervised trained segmentation models may be not mandatory for projects when a median DSC of 0.82 is sufficient. 

Keywords: AI/ML Image Reconstruction, Machine Learning/Artificial Intelligence

Motivation: The synthetic training data used for mapping reconstruction of deep learning can simulate the required features, but it is difficult to fully simulate the unnecessary characteristics existing in real-world data.

Goal(s): This work aims to enable the model to extract the essential mapping relationships for mapping reconstruction and eliminate the interference of non-ideal factors in real data.

Approach: We propose a mask pre-training method called Masked U-net that allows the model to learn appropriate inductive biases on quantitative images.

Results: The proposed method can better extract relevant features and reduce the interference of irrelevant factors in real data.

Impact: The proposed method bridges the gap between the real data and synthetic data, improves the quality of deep learning reconstruction driven by synthetic training data, and achieves important application in T₂/T₂^* mapping reconstruction of multiple overlapping-echo detachment (MOLED) imaging.

Keywords: Head & Neck/ENT, Head & Neck/ENT

Motivation: Distortion-free brain diffusion-weighted imaging (DWI) remains a challenge due to susceptibility artifacts and low signal-to-noise ratio (SNR).

Goal(s): Assess the effectiveness of Compressed SENSE (CS) combined with deep learning (CS-DL) in improving brain DWI image quality. 

Approach: We compared various acceleration schemes and reconstruction methods on TSE-DWI brain images.

Results: CS-DL with a factor of 4 improved image quality and SNR, while reducing scan time by 22%. 

Impact: Implementation of CS-DL in TSE-DWI holds promise for high-fidelity, distortion-free imaging, facilitating detailed analysis of small brain abnormalities in regions affected by magnetic field inhomogeneity and susceptibility.