Oral Abstract Session
SCMR 22nd Annual Scientific Sessions
Ahmed Fahmy, PhD, MSc
Research Fellow
Beth Israel Deaconess Medical Center and Harvard Medical School
Marcel Beetz, MSc
Researcher
Technical University of Munich
Ulf Neisius, MD, PhD
Research Fellow
Beth Israel Deaconess Medical Center and Harvard Medical School
Raymond Chan, MD
MD
Toronto General Hospital
Martin Maron, MD
Assistant Professor
Tufts Medical Center
Evan Applebaum, MD
Physician
Men's Health Boston
Bjoern Menze, PhD
Assistant Professor
Technical University of Munich
Reza Nezafat, PhD
Associate Professor of Medicine
Department of Medicine (Cardiovascular Division) Beth Israel Deaconess Medical Center
Background:
Scar quantification in late gadolinium enhancement (LGE) is often performed using manual or semi-automatic techniques based on signal thresholding. However, currently there is no gold-standard approach for scar quantification and scar volume is highly dependent on threshold value used in these techniques. We recently proposed a deep fully convolutional neural network (CNN) method for automatic quantification of myocardial scar and left ventricular (LV) volumes for patients with hypertrophic cardiomyopathy (HCM) (1). The developed CNN method was limited to the segmentation of a single 2D slice at a time and thus was potentially limited by missing important 3D information from neighboring slices. In this study, we sought to develop a 3D CNN model that incorporates data from adjacent slices for scar quantification.
Methods: A 3D deep CNN based on the Unet architecture was trained and tested using 7775 manually-segmented short axis LGE images from 1041 HCM patients. The image set, acquired and analyzed as a part of multicenter clinical study (2), was split at random into training (852 patients) and testing (189 patients) subsets such that balanced number of images from different centers and different scar sizes was maintained in each subset. The input and output of the network were fixed to 3 slices of consecutive LGE images and labeled images, respectively. For patient datasets with more than 3 slices, a moving window with step of one slice was used to process every consecutive three slices (Fig.1). Due to the overlapping of the moving windows, some slices were processed more than once and thus an image fusion (by taking the union of the segmented regions in the different images) was used to yield one labeled image per slice. The performance of the automatic quantification of scar volume and LV mass versus manual quantification was evaluated using correlation analysis on per patient and per slice basis.
Results:
Fast segmentation of LGE images was achieved in < 0.3 sec/image (Fig. 2). The automatically segmented scar showed strong correlation with the manually segmented scar in both per-patient (r = 0.85; slope=0.90; p < 0.001) and per-slice (r = 0.87; slope=0.96; p < 0.001) analyses. Automatic segmentation of LV showed strong correlation with manual segmentation in per-patient (r = 0.97; slope=1.16; p < 0.001) and per-slice (r = 0.97; slope=1.05; p < 0.001) comparisons.
Conclusion: Three-dimensional deep neural network based segmentation of 3D LGE images allows fast and accurate quantification of LV mass and scar in HCM patients.