QF1-009 - Automated Left Ventricle Segmentation of Cardiac LGE with Deep Learning

Thursday, February 13, 2020

3:09 PM – 3:12 PM

Location: Sabal/Sago (Meeting Room in Exhibit Hall)

Background: Late Gadolinium Enhancement (LGE) cardiovascular magnetic resonance (CMR) is the gold standard test for assessment of myocardial scarring (1). While scar volume quantification is potentially a useful predictor of major adverse cardiac event (MACE), it is rarely used in clinical practice due to lengthy image segmentation time. In this study, we sought to develop a fully automated approach for LV scar quantification using deep learning (DL).

Methods: This study leveraged our access to labeled LGE data with myocardial wall and scar contours from the Biologic Markers and MRI SCD Cohort Study. As shown in Figure 1, we implemented a 4-class residual U-Net (2) on a GPU workstation (NVIDIA Tesla V100) equipped with Pytorch, in order to automatically segment the whole left ventricle into 4 classes: background, healthy myocardium, scar, and LV blood pool. Labeled data were randomly selected from 52 patients (482 2D images in total) were used for training the network, and the remaining labeled data from 10 patients (94 2D images in total) were used for testing. We used Pytorch built-in cross entropy loss combined with dice loss of the whole LV (i.e. Myocardium outer boundary). We compared our approach to a previously described method (3) using 3 classes (background, healthy myocardium and scar), where both our approach and previous approach used the same manual contour files, as shown in Figure 2. To evaluate the network performance, we calculated Sørensen–Dice index and scar volume for each testing subject with manual contours as reference.

Results: Figure 3 shows results from 4 representative patients, in which 4-class was slightly better than 3-class. As summarized in Table 1, the mean DICE of 4-class U-Net for the whole LV was 0.957±0.008. The mean DICE scores for myocardium and scar were not significantly different (P >0.05) between 4-class than 3-class. The scar volume was not significantly (P >0.05) different between manual, 3-class U-Net, and 4-class U-net using 1-way ANOVA.

Conclusion: Our proposed 4-class U-Net showed the potential of automatic segmentation of the whole LV for LGE images, which is also more robust than 3-class U-Net model. The extra labeling of LV blood pool improved the performance of deep learning network. Future investigation will include more datasets for training and testing, as well as incorporation of models to constrain/refine learning.

Presenting Author(s)

Daming Shen, BSc

PhD candidate
Northwestern University

Disclosure: Disclosure information not submitted.

Primary Author(s)

Daming Shen, BSc

PhD candidate
Northwestern University

Disclosure: Disclosure information not submitted.

Co-Author(s)

Daniel Kim, PhD

Professor
Northwestern University

Disclosure: I do not have any relevant financial / non-financial relationships with any proprietary interests.
BB

Brandon C. Benefield, MSc

Research Lab Manager
Northwestern University

Disclosure: Disclosure information not submitted.
JB

Justin Baraboo, BSc

Graduate Student
Northwestern University

Disclosure: Disclosure information not submitted.
Michael Markl, PhD

Professor & Vice Chair for Research
Northwestern University

Disclosure: Circle Cardiovascular Imaging: Consulting Fees (e.g., advisory boards), Grant/Research Support; Cryolife Inc (Individual(s) Involved: Self, Products/Services: grant): Grant/Research Support; NXT Biomedical (Individual(s) Involved: Self, Products/Services: Consultant): Consulting Fees (e.g., advisory boards); Siemens Healthcare: Grant/Research Support
Daniel C. Lee, MD, MSc

Associate Professor of Medicine (Cardiology) and Radiology
Northwestern Memorial Hospital

Disclosure: I do not have any relevant financial / non-financial relationships with any proprietary interests.