Cardiovascular MR Center at Harvard

DL-Rad-ScarPrediction

Code for identifying patients without scar using DL and Radiomics analyses of non-Gd bSSFP cine sequences.

DRAPR

We implemented a 3D (2D+time) convolutional neural network to suppress streaking artifacts from undersampled radial cine images. We trained the network using synthetic real-time radial cine images simulated using ECG-gated segmented Cartesian k-space data, which was acquired from 503 patients during breath-hold and at rest. Further, we implemented a prototype real-time radial sequence with acceleration rate = 12 on a 3T scanner, and used it to collect cine images with inline DL reconstruction whose total reconstruction time was 16.6 ms per frame. We evaluated the performance of the proposed approach by initially recruiting 9 healthy subjects in whom only rest images were collected. Subsequently, we recruited 14 subjects who participated in an exercise CMR imaging protocol in which both rest and post-exercise images were collected, including 8 patients with suspected coronary artery disease. Exercise was done using a CMR-compatible supine cycle ergometer positioned on the MR table.

ImageSharpenessMeasurement

The repository contains a source code for measuring image sharpness. Image sharpness is defined as the absolute gradient of intensity profile extracted from the normalized image.

ImprovingMyoMapNetT1Estimation

Myocardial_Scar_Detection

Code for the paper "Gadolinium-Free Cardiac MRI Myocardial Scar Detection by 4D Convolution Factorization" MICCAI2023

MyoMapNet

We implemented a FC that uses pixel-wise T1-weighted signals and corresponding inversion time to estimate T1 values from a limited number of T1-weighted images. we studied how training the model using native, post-contrast T1 and a combination of both could impact performance of the MyoMapNet. We also explored two choices of number of T1 weighted images of four and five for native T1, selected to allow training of network using existing data from modified Look-Locker sequences (MOLLI). After a rigorous training using in-vivo T1 maps of 607 patients, undergoing clinical cardiac MR exams, collected by MOLLI, the performance of MyoMapNet was evaluated using in-vivo data of 61 patients by discarding the additional T1-weighted images from MOLLI. Subsequently, we implemented LL4 T1 mapping sequence and an inline implementation of MyoMapNet on a 3T Siemens scanner to imaging and inline reconstruction of T1 maps. The inline MyoMapNet was then used to collect LL4 T1 and MOLLI in 16 subjects to demonstrate feasibility of inline MyoMapNet.

MyoMapUnet

We implemented and tested three different classes of deep learning architectures for MyoMapNet: (a) a fully connected neural network (FC), (b) convolutional neural networks (VGG19, ResNet50), and (c) encoder-decoder networks with skip connections (ResUNet, U-Net). Modified Look-Locker Inversion Recovery (MOLLI) images from 749 patients at 3T were used for training, validation, and testing. The first four T1 weighted images from MOLLI5(3)3 and/or MOLLI4(1)3(1)2 protocols data were extracted to create accelerated cardiac T1 mapping data. We also prospectively collected data from 28 subjects using MOLLI and LL4 to further evaluate model performance.

RealTimeTaggingSSFP

REGAIN

The repository contains a source code for Resolution-Enhancement-Generative-Adversarial-Inline-Network (REGAIN). REGAIN is a generative adversarial neural network for REGAINing image sharpness and Spatial resolution. The trained network generates a resolution-enhanced image in the Phase-encoding (ky) direction in MRI.

SimulationPhysicsMR

HMS-CardiacMR/MyoMapNet

We implemented a FC that uses pixel-wise T1-weighted signals and corresponding inversion time to estimate T1 values from a limited number of T1-weighted images. we studied how training the model using native, post-contrast T1 and a combination of both could impact performance of the MyoMapNet. We also explored two choices of number of T1 weighted images of four and five for native T1, selected to allow training of network using existing data from modified Look-Locker sequences (MOLLI). After a rigorous training using in-vivo T1 maps of 607 patients, undergoing clinical cardiac MR exams, collected by MOLLI, the performance of MyoMapNet was evaluated using in-vivo data of 61 patients by discarding the additional T1-weighted images from MOLLI. Subsequently, we implemented LL4 T1 mapping sequence and an inline implementation of MyoMapNet on a 3T Siemens scanner to imaging and inline reconstruction of T1 maps. The inline MyoMapNet was then used to collect LL4 T1 and MOLLI in 16 subjects to demonstrate feasibility of inline MyoMapNet.

HMS-CardiacMR/DRAPR

We implemented a 3D (2D+time) convolutional neural network to suppress streaking artifacts from undersampled radial cine images. We trained the network using synthetic real-time radial cine images simulated using ECG-gated segmented Cartesian k-space data, which was acquired from 503 patients during breath-hold and at rest. Further, we implemented a prototype real-time radial sequence with acceleration rate = 12 on a 3T scanner, and used it to collect cine images with inline DL reconstruction whose total reconstruction time was 16.6 ms per frame. We evaluated the performance of the proposed approach by initially recruiting 9 healthy subjects in whom only rest images were collected. Subsequently, we recruited 14 subjects who participated in an exercise CMR imaging protocol in which both rest and post-exercise images were collected, including 8 patients with suspected coronary artery disease. Exercise was done using a CMR-compatible supine cycle ergometer positioned on the MR table.

HMS-CardiacMR/ImageSharpenessMeasurement

The repository contains a source code for measuring image sharpness. Image sharpness is defined as the absolute gradient of intensity profile extracted from the normalized image.

HMS-CardiacMR/Myocardial_Scar_Detection

Code for the paper "Gadolinium-Free Cardiac MRI Myocardial Scar Detection by 4D Convolution Factorization" MICCAI2023

HMS-CardiacMR/ImprovingMyoMapNetT1Estimation

HMS-CardiacMR/MyoMapUnet

We implemented and tested three different classes of deep learning architectures for MyoMapNet: (a) a fully connected neural network (FC), (b) convolutional neural networks (VGG19, ResNet50), and (c) encoder-decoder networks with skip connections (ResUNet, U-Net). Modified Look-Locker Inversion Recovery (MOLLI) images from 749 patients at 3T were used for training, validation, and testing. The first four T1 weighted images from MOLLI5(3)3 and/or MOLLI4(1)3(1)2 protocols data were extracted to create accelerated cardiac T1 mapping data. We also prospectively collected data from 28 subjects using MOLLI and LL4 to further evaluate model performance.

HMS-CardiacMR/DL-Rad-ScarPrediction

Code for identifying patients without scar using DL and Radiomics analyses of non-Gd bSSFP cine sequences.

HMS-CardiacMR/REGAIN

The repository contains a source code for Resolution-Enhancement-Generative-Adversarial-Inline-Network (REGAIN). REGAIN is a generative adversarial neural network for REGAINing image sharpness and Spatial resolution. The trained network generates a resolution-enhanced image in the Phase-encoding (ky) direction in MRI.

HMS-CardiacMR/DENT

We developed a highly accelerated high-frame-rate cine for Ex-CMR by accelerating spatial resolution using REGAIN, followed by synthesizing new frames using Deformation ENcoding Transformer (DENT).

HMS-CardiacMR/Radiomics_Histology_NIDCM

HMS-CardiacMR/RealTimeTaggingSSFP

HMS-CardiacMR/Scar_enhancement

Myocardial Scar Enhancement in LGE Cardiac MRI using Localized Diffusion

HMS-CardiacMR/SimulationPhysicsMR

HMS-CardiacMR/CineROI

An open-source, flexible, plug-and-play inline CMR segmentation platform for the rapid deployment of ML models into the clinical workflow.

HMS-CardiacMR/CRISPFlow

Accelerated Phase Contrast MRI with Use of Resolution Enhancement Generative Adversarial Neural Network

HMS-CardiacMR/FastCSE

Accelerated Chemical Shift Encoded Cardiac MRI with Use of Resolution Enhancement Network

HMS-CardiacMR/ML-Predtiction-NICM

Code for Fahmy et al. "An Explainable Machine Learning Approach Reveals Prognostic Significance of Right Ventricular Dysfunction in Nonischemic Cardiomyopathy". JACC Cardiovasc Imaging. 2022 doi: 10.1016/j.jcmg.2021.11.029.

HMS-CardiacMR/Multicenter_MyoMapNet

HMS-CardiacMR/NICM_ICD_Radiomics

HMS-CardiacMR/Radiomics-SCD-HCM

Python code used to investigate the prognostic value of Radiomics analysis of LGE in predicting sudden cardiac death (SCD) in HCM patients..

HMS-CardiacMR/Radiomics_Histology_PCA