Machine Learning and Artificial Intelligence

1350612 - Myocardial Segmentation of Tagged Magnetic Resonance Images with Deep Learning

Wednesday, January 25, 2023

Location: E-POSTER

Presenting Author(s)

ML

Michael Loecher, PhD

Research Scientist
Stanford University, United States

Primary Author(s)

AD

Arnaud P. Dhaene, MSc

Visiting Student Researcher
Stanford University
Lausanne, Vaud, Switzerland

Co-Author(s)

ML

Michael Loecher, PhD

Research Scientist
Stanford University, United States
AW

Alexander J. Wilson, PhD

Postdoctoral Scholar
Stanford University, United States
Daniel B. Ennis, PhD

Professor
Stanford University
Stanford, California, United States

Disclosure(s):

Daniel B. Ennis, PhD: No relevant disclosure to display

Background: The use of deep learning (DL) segmentation in cardiac MRI applications seeks to improve the accessibility and efficiency of cardiovascular functional quantification and classification as it can improve diagnosis and treatment planning for related pathologies [1]. While high-performing myocardial segmentation models exist for cine images [2], this is not the case for tagged images. Such segmentation models are essential for the correct initialization of motion tracking algorithms, integral to automated myocardial strain estimation [3].

Methods: Fully segmented (labeled) tagged images are scarce due to clinical resource constraints and time-intensive annotation procedures. Herein, common biomedical segmentation network architectures, nnUnet and ResNetVAE, were pretrained with public datasets of cine images, transformed into their tagged-appearing equivalents following a physics-driven and a distinct generative CycleGAN approach [4]. Subsequently, models were fine-tuned on a subset of real tagged images, annotated by an expert researcher. In conjunction, the effect of a boundary-aware loss function on training was evaluated via grid-search. Models were evaluated using the Dice coefficient (DSC) and the 95th-percentile Hausdorff distance (HD-95).

Results: Pretrained models performed better and converged faster than models trained from scratch as they were seen to have a higher average DSC by up to 2.7% and a lower average HD-95 by up to 1.4 mm. While all segmentations presented a disc-like shape, the masks produced by ResNetVAE were physiologically more consistent than the ones produces by nnUnet. The ResNetVAE architecture outperformed the nnUnet’s by up to 4.3 % DSC and 1.10 mm HD-95, reaching a DSC of 0.828 and HD-95 of 4.745 mm on a held-out test set. Moreover, nnUnet models were less robust to varying patient health conditions as they perform better on images from healthy subjects in comparison to DMD patients. The DSC distribution was also higher for images from the first 15 to 30% of timeframes and worsens as the taglines fade during the cardiac cycle. The optimal contribution factor for the boundary-aware distance loss was 0.05, improving average performance by up to 3.9% DSC and 2.25 mm HD-95.

Conclusion:
While the nnUnet architecture is more commonly used in biomedical segmentation applications, when trained on tagged CMR images the smaller ResNetVAE performed better, produced more physiologically consistent masks, and was more robust to varying patient health conditions. Despite variations in performance through acquisition timeframes, the developed model sets a standard for myocardium segmentation in tagged CMR images, allowing for the automation of an important task within the myocardial strain estimation pipeline.

Acknowledgement
This work was partly supported by NIH R01 HL152256.