SCMR 23rd Annual Scientific Sessions

Techniques and Methods Track

Techniques and Methods Session 5: Quantitative CMR for Tissue Characterization

5 - Automated Analysis of Myocardial Pre/Post T1 and ECV Using Deep Learning

Saturday, February 15, 2020

10:35 AM – 10:50 AM

Location: Salon J1

Background: Delineation of endo/epicardial boundaries is needed for T1 mapping analysis. Manual contouring is laborious and subjective. Recently, automated methods for Native T1 map segmentation were presented, some using deep learning [1,2]. Here we present a fully automatic deep learning framework for performing pre/post Gd contrast T1map segmentation to enable calculation of partition-coefficient (λ) and ECV values based on AHA segments.

Methods: A total of 90 patients were scanned across 4 clinical categories: Normal and 3 different cardiomyopathies (Ischemic, Non-Ischemic and Hypertrophic). Stratified random sampling was used to select patients: 48 for training, 12 for validation, 30 for testing. A 11 h.b MOLLI sequence was used to obtain 3 short axis slices for native and post-Gd contrast. After registration using MIND [3], a T1 map and a synthetic image from a single inversion time (TI) were reconstructed. For training, an expert drew myocardial contours on a single MOLLI image on both native and post-Gd images. A standard U-Net architecture pre-trained on CINE images was used with transfer learning [4,5]. Three different inputs were compared for their performance with the same contours: MOLLI image, T1 (or R1) map and the synthetic image. R1map was used for post-Gd images. Data-augmentation was done using limited affine transformation, orientation and TI-based contrast adjustments. A single network was used to train both native and post-Gd images. The network outputs were further processed for analysis of segmental T1 and λ. Two different experts provided contours for the test studies. The Dice metric on the myocardial tissue was used for pair-wise comparison between the experts and network predictions.

Results: Validation results: Training and prediction on MOLLI images themselves resulted in the best performance (Myocardial Dice = 0.86 ± 0.05). Next best was on Synthetic images with contrast augmentation (Dice = 0.85 ± 0.05). Training on T1 maps was least performing (Dice = 0.81 ± 0.06). We selected the synthetic images for testing since it is independent of the acquisition sequence and performed better than T1 maps.
Test results: After excluding two failures, the network performed similarly when compared to the two experts (Dice=0.78 and 0.79, respectively). This was very close to the inter-observer Dice of 0.78.

Conclusion: This is the first deep learning network that demonstrates automated myocardial segmentation of both pre and post-Gd T1 maps, thus allowing for ECV analysis AHA-segment wise. It also has other advantages. Due to the use of synthetic images, it is compatible with several MOLLI acquisition schemes and with other T1 mapping sequences as well. It requires less effort to train: the expert draws only on a single TI image and it uses transfer learning for faster convergence. The performance of our network is promising - the validation Dice is similar to prior findings [1] and the test Dice shows human level performance as reflected by the inter-observer variability.

Presenting Author(s)

VR

Venkat Ramanan

Research Associate
Sunnybrook Research Institute

Disclosure: Disclosure information not submitted.

Primary Author(s)

VR

Venkat Ramanan

Research Associate
Sunnybrook Research Institute

Disclosure: Disclosure information not submitted.

Co-Author(s)

VR

Venkat Ramanan

Research Associate
Sunnybrook Research Institute

Disclosure: Disclosure information not submitted.
NB

Nitish Bhatt

University of Waterloo

Disclosure: Disclosure information not submitted.
XQ

Xiuling Qi, PhD

Research Associate
Sunnybrook Research Institute

Disclosure: Disclosure information not submitted.
TY

Ting Yu, BSc

McMaster University

Disclosure: Disclosure information not submitted.
MN

Matthew Ng, BSc

Sunnybrook Research Institute

Disclosure: Disclosure information not submitted.
IR

Idan Roifman, MD, MSc

Staff Cardiologist
Sunnybrook Health Sciences Centre

Disclosure: Disclosure information not submitted.
LG

Lancia Guo, MD

Sunnybrook Health Sciences Centre

Disclosure: Disclosure information not submitted.
LJ

Laura Jimenez-Juan, MD

Physician
Sunnybrook Health Sciences Centre

Disclosure: Disclosure information not submitted.
Graham A. Wright, PhD

Sr. Scientist
Sunnybrook Research Institute

Disclosure: Disclosure information not submitted.
NG

Nilesh R. Ghugre, PhD

Scientist
Sunnybrook Research Institute / University of Toronto

Disclosure: Disclosure information not submitted.