Machine Learning and Artificial Intelligence

1352475 - Adaptive Weighted Categorical Cross Entropy - A New Loss Function for Automatic Segmentation of Left Ventricular Myocardium and Scar from 3D LGE-CMR

Wednesday, January 25, 2023

Location: E-POSTER

Presenting Author(s)

Vivek P. Jani, MS

MD-PhD Student
The Johns Hopkins University
Baltimore, Maryland, United States

Primary Author(s)

Vivek P. Jani, MS

MD-PhD Student
The Johns Hopkins University
Baltimore, Maryland, United States

Co-Author(s)

WL

Wei Hao Lee, BA

Cardiac MRI Data Analyst
Johns Hopkins Medicine, United States
MO

Mohammad Ostovaneh, MD

Cardiology Fellow
The John Hopkins Hospital, United States
EC

Ela Chamera

Research Sonographer/Imager
The Johns Hopkins University
baltimore, Maryland, United States
Yoko Kato, MD, PhD

Research Fellow
The Johns Hopkins University
Baltimore, Maryland, United States
SK

Shelby Kutty, MD, PhD

Helen B. Taussig Professor
The Johns Hopkins University, Nebraska, United States
Joao A. C Lima, MD

MD
The John Hopkins Hospital
Baltimore, Maryland, United States
Bharath Ambale-Venkatesh, PhD

Physicist
The John Hopkins Hospital
Baltimore, Maryland, United States

Background: Widespread clinical adoption of scar quantification is hindered by the significant time associated with analysis. Here, we sought to train convolutional neural networks for automatic segmentation of ventricular myocardium and scar from a large database of ischemic late gadolinium enhanced (LGE) scar images from across 32 different sites. We developed a new training paradigm, denoted as adaptive weighted categorical cross entropy, a new loss function that that decreases the weights associated with non-background voxels with increasing epochs.

Methods:

This study included 156 3D LGE short axis images of the myocardium. A standard 3D U-Net was trained for segmentation of myocardium and scar. All convolution kernels included batch normalization, kernel regularization, and spatial dropout to limit overfitting. All models were trained on 20,000 augmented images from 117 unique studies. Weights for U-Net were first initialized with a weighted categorical cross entropy loss function, which penalizes pixels that were misclassified as background during training. Following weight initialization, weights were decayed on a schedule as epochs increased, eventually converging to a standard categorical cross entropy loss function. All models were evaluated on an internal validation set (n=39) and external test set (n=33). Train and test set accuracy were quantified utilizing a dice coefficient.

Results:

With adaptive weighted cross entropy, models were all found to converge after approximately 10 epochs. During training, we observed that models learned the left ventricular segmentation maps early and identified scar voxels in later epochs. Median overall dice coefficient was 0.63 on the internal validation set (maximum: 0.85). Concordance correlation coefficient for overall scar percent was 0.56 on the internal test set and 0.50 on the external test set. Bland-Altman analysis and correlations for scar percent in the internal and external test sets are shown in Figure 1. Accuracy is low for scars < 10% on the external test set, where the deep learning model overestimates scar. This is consistent with the low percentage of cases (11/117) with < 10% scar in the training set. Bland-Altman analysis of segmental scar revealed minimal bias for percent basal (5±8%) and mid-ventricular (5±7%) scar. Representative deep learning segmentation is shown in Figure 2.

Conclusion: Adaptive weighted cross entropy allows for efficient convergence of 3D U-Nets for ventricular and scar segmentation despite class imbalance.