Synopsis

We present a fully automated approach for quantification of left ventricle scar in Late Gadolinium Enhancement (LGE) cardiac MR (CMR), using a residual neural network. LGE images were acquired in 1075 patients with known hypertrophic cardiomyopathy in a multi-center clinical trial. Scar segmentation was performed in all patients by a CMR-trained cardiologist. For training, we use a two-phase procedure, using cropped and full-sized images consecutively. We train different models using sigmoid cross-entropy loss and Dice loss and measure average LV segmentation Dice scores of 0.77 ± 0.10 and 0.70 ± 0.12 and estimated scar percentage mismatches of 3.59% and 3.00%, respectively.

Introduction

Methods

We propose to use a residual neural network architecture to automatically segment LV scar. A database consisting of 2D LGE images from 1075 patients with known hypertrophic cardiomyopathy acquired as part of a multi-center clinical trial was used to evaluate the performance of the approach.⁵ An experienced cardiologist with 5 years of experience in CMR manually segmented all images, which were used as “ground truth”.

A common problem in the segmentation of small objects, such as scar, is the high imbalance between background and foreground pixels, which is often tackled by elaborate loss-weighting schemes.⁶ We circumvent this challenge by applying a residual neural network architecture,⁷ which has shown competitive performance without weighted loss, when adapted for semantic segmentation.⁸ We set up our architecture (Figure 1) largely following these adaptions. Upsampling is realized by backwards-strided convolutions initialized for bilinear interpolation⁹ and dropout of 0.5 is applied in bottleneck layers during training.

We define the softmax cross-entropy loss

$$L_{x} = -\sum_{i} log(s_{i}^{c_{y}})\,,$$

using softmax probabilities $$$s_{i}^{c} \in [0,1]$$$ for classes $$$c \in \{0, 1, 2\}$$$ (background, healthy LV, scar), with $$$s_{i}^{c_{y}}$$$ being the softmax probability for the correct class at pixel $$$i$$$.

We define a second loss function, based on the Dice coefficient for the predicted segmentations:

$$L_{d} = \frac{1}{n}\sum_{c}(1 - \frac{2\sum_{i} s_{i}^{c} y_{i}^{c}}{\sum_{i} s_{i}^{c} + \sum_{i} y_{i}^{c}})\,.$$

Here, we calculate the average Dice loss for all $$$n$$$ classes, using softmax probabilities $$$s_{i}^{c}$$$ and ground truth label $$$y_i^{c}\in\{0,1\}$$$ for class $$$c$$$ at pixel $$$i$$$. Optimization is performed using the Adam method with learning rate of 0.001 and exponential decay rates $$$\beta_1$$$=1, $$$\beta_2$$$ = 0.999.¹⁰

The LGE images first undergo several preprocessing steps. Volumes are resampled to a pixel spacing of 1.2 mm and padded or cropped to a size of 256 x 256 pixels in the xy-plane. We augment training data by performing random translation, mirroring and elastic deformation on volumes during training.⁶ 2D slices are used as input, after being normalized to zero mean and unit variance. We split the data into training and test sets, each containing 80% and 20% of the total data, respectively.

We follow a two-phase procedure for the training of the models (see Figure 1). First, the network is pre-trained with cropped, 128 x 128 sized patches that are extracted along the short axis of the heart. We train the network for up to 80000 iterations with batch size of 20. The best performing model is then fine-tuned for up to 20000 iterations with batch size 10 on full-sized images. By pre-training with cropped images, we reduce training time, while implicitly reducing the imbalance between background and foreground pixels. At test time, we apply a rectangular ROI filter on each volume to remove potential outliers.

We perform quantitative evaluation on models generated with the different loss functions. In addition to the volume-wise LV Dice coefficient, we measure the accuracy of the predicted volume-wise scar percentage and pixel quantity. We evaluate two models trained with cross-entropy loss with best performance in LV segmentation and scar percentage prediction, respectively. Furthermore, we evaluate the model with best overall performance trained with Dice loss as well as an ensemble of all three models.

Results

Conclusion and Discussion

Acknowledgements

References