Keywords: Analysis/Processing, Spinal Cord, Localizer images, MRI, Spine, Segmentation, Deep Learning

Motivation: Obtaining consistent spine MRI images irrespective of patient posture, spine deformities, and technologists’ skills, with minimal disruption in the existing workflow.
 

Goal(s): To develop an intelligent scan plane prescription for spine MRI using deep learning on regular 3-plane localizer images.
 

Approach: We adopted a multi-resolution CNN network for multiple segmentation tasks - spine vertebrae, intervertebral disc (IVD), and saturation band (SB) across all the spine stations (cervical, thoracic, and lumbar) and orientations (sagittal and coronal).

Results: We reported good segmentation of vertebrae and IVD, along with consistent SB placement with angle error of less than 5 degree and no overlap with the spine region.

Impact: We present a first-of-its-kind integrated multi-label 3D DL model that operates on 2D 3-plane regular localizers to aid consistent MRI scan planning. This model combines MRI localizer images across orientation, across spine stations, and across multiple imaging tasks.

Keywords: Analysis/Processing, Brain, automatic scan planning

Motivation: Manual scan planning in clinical MRI is inaccurate, inconsistent and time-consuming.

Goal(s): A deep learning-based end-to-end automated scan planning framework has been developed for MRI head scans.

Approach: We propose a two-stage end-to-end 3D cascaded convolutional network framework, called 3D CFP-UNet, which localizes the positions of five key anatomical landmarks and achieves a coarse-to-fine result. We also propose loss functions PRL and DRL with physical meaning in automatic scan planning. 

Results: Our approach yields satisfactory scan planning results on 229 test subjects, with PAE and PRE reaching 0.872mm and 0.10%, respectively.

Impact: MRI automated scan planning can help improve scan efficiency. Also, it improves scan consistency for follow-up comparisons.

Keywords: Analysis/Processing, Microstructure, Histology, Diffusion Imaging, g-ratio, axon diameter

Motivation: Machine learning approaches are an alternative to conventional biophysical model fitting used to generate MRI microstructural maps, but the lack of paired MRI-histology data complicates end-to-end training of these models.

Goal(s): Develop a nonparametric deep learning based prediction of joint distributions of g-ratios and axon diameters from multimodal MRI data.

Approach: Histology-based synthetic MRI data was used to pretrain a conditioned normalizing flow model. Transfer learning was then performed on limited paired MRI-histology data.

Results: The joint distribution shows good visual agreement with actual samples and the distances between the marginal probabilities and their respective samples exhibit a Jensen-Shannon distance smaller than 0.22.

Impact: We present an optimized model to obtain non-parametric joint distributions of g-ratios and axon diameters from multimodal MRI from limited experimental data. The approach can easily be adapted to other microstructural modeling tasks.

Keywords: Analysis/Processing, Spectroscopy, Brain, Deep learning, Quantification, RNN

Motivation:
Incomplete FIDs can be obtained due to the limited sampling windows as in spectroscopic MRF and SSFP-MRSI, or due to FID truncation for removing spectral artifact.

Goal(s): Developing a means of quantifying metabolites from incomplete FIDs will allow more efficient sequence design and better experimental outcome.

Approach: We developed a recurrent-neural-network (RNN) for metabolite quantification from incomplete FIDs at 3.0T. The RNN was trained on simulated data and tested on in vivo data.

Results: Although the performance of the RNN requires further improvement for low concentration metabolites (e.g., GABA), it may allow quantification of the major metabolites under highly limited sampling windows.

Impact: Incomplete FIDs can be obtained due to the limited sampling windows as in spectroscopic MRF and SSFP-MRSI. We developed a recurrent-neural-network, which can quantify the major metabolites from the initial 64 FID data points, thereby allowing more efficient sequence design.  

Keywords: Analysis/Processing, Radiomics

Motivation: To investigate the effect of a deep learning reconstruction algorithm on radiomic image features.

Goal(s): To assess the effect of AIR^TM Recon Deep Learning (ARDL), a commercial AI reconstruction algorithm, on radiomic features in a set of phantoms.

Approach: A set of radiomic phantoms were constructed and used to acquire images with different numbers of signal averages and ARDL levels. Effects were evaluated through intraclass correlation coefficient (ICC) measures.

Results: Radiomic features maintain excellent ICC values (>0.9) at a constant SNR with ARDL Low, but ICC values decrease with higher ARDL levels

Impact: This research highlights how deep learning image reconstruction can alter radiomic features and could help define a subset of stable features. The level of deep learning reconstruction applied is shown to have significant impact, even at constant SNR.

Keywords: Analysis/Processing, Fat, Proton Density Fat Fraction

Motivation: Most Deep Learning (DL) methods to estimate liver PDFF require reference results for training and can only calculate deterministic outputs with unknown uncertainty.

Goal(s): To estimate liver PDFF using a fully-unsupervised DL method for MR water-fat separation capable of quantifying uncertainty.

Approach: We propose a physics informed DL-based framework which can be trained purely on chemical shift-encoded MR images. Our method estimates stochastic R2* and Δf maps, enabling uncertainty quantification, which are then used to obtain stochastic water-only and fat-only components.

Results: Liver PDFF estimations showed good agreement with a reference technique, and uncertainty maps associated with imperfections in the considered physical model.

Impact: The proposed physics-informed DL model requires only MR data for training, which facilitates the data gathering process. Moreover, our uncertainty-aware approach can quantify the uncertainty associated to the final estimations, which may be of significant value in clinical practice.

Keywords: Analysis/Processing, Prostate, Quality Control, DWI, ADC, Preduiction

Motivation: ADC maps are an essential tool for early prostate cancer detection but are often uninterpretable due to imaging artifacts

Goal(s): Detect problems early in the imaging procedure using T2 images to predict the future quality of the ADC map

Approach: Constructed a multisite corpus of 486 patients imaged at both the NIH and outside. Investigated the influence of acquisition parameters on image quality and the predictive power of neural networks and simple anatomy measurements from the T2 image

Results: ADC image quality can be predicted from the T2 image using either a neural network approach or measurement of the rectal cross-section

Impact: The probability of a low quality, uninterpretable ADC maps can be inferred early in the imaging process, allowing corrective action (e.g. removal of gas by a muscle relaxant) to be employed

Keywords: Analysis/Processing, Machine Learning/Artificial Intelligence

Motivation: Computational fluid dynamics (CFD) is used for non-invasive cardiovascular hemodynamic assessment, but it is limited by time-consuming manual segmentation and expertise needed for simulation.

Goal(s): Improve the speed and simplify volume mesh generation and CFD flow field calculation.

Approach: Develop a single deep-learning model capable of reconstructing the pulmonary artery from a 3D cardiac MRI as a volume mesh and predicting CFD-like pressure and flow.

Results: Our model achieves accurate pulmonary artery reconstruction with a median Dice score of 0.9. It computes CFD-like pressure and flow with median errors of 14.9% and 9.0%, respectively. Our model is ~10,000 times faster than manual calculation.

Impact: Image2Flow is a single-pass deep-learning model that rapidly and accurately reconstructs pulmonary artery volume meshes from 3D cardiac MR and predicts CFD-like flow fields. Our model can potentially streamline and expedite cardiovascular haemodynamic assessment and facilitate more efficient treatment planning.

Keywords: Analysis/Processing, Segmentation, Cortical Surface Reconstruction

Motivation: Cortical surface reconstruction (CSR) is important for surface-based analysis of the structure and function of the cerebral cortex.

Goal(s): We present an efficient method for simultaneous CSR and spherical mapping, all within a matter of seconds. Inherent correspondence allows easy and direct mapping of geometric features from cortical surfaces to the sphere.

Approach: Our flow-based method learns velocity fields to deform a spherical template mesh to the cortical surfaces with one-to-one vertex correspondence for direct spherical mapping.

Results: Using data from the Baby Connectome Project (BCP), we demonstrate that our method predicts more accurate and uniform surface meshes compared with several state-of-the-art methods.

Impact: Our method provides a way for fast and accurate infant cortical surface reconstruction. The one-to-one vertex correspondence between template sphere and the cortical surfaces enables easy and direct downstream analyses.