gms | German Medical Science

Introduction: Myocardial infarction is the number one of the most common heart diseases. The decreased blood flow causes damaged and scarred heart tissue. Healthy tissue compensates the damaged tissue, causing structural changes of the left ventricle.

For the evaluation of the effects of the myocardial infarct, magnetic resonance images (MRI) are taken. Quantitative measurements, e.g. volume and mass, based on the boundaries of the outer layer (epicardium) and the inner layer (endocardium), are used to determine the state of the left ventricle.

Neural networks, especially convolutional neural networks (CNNs) achieved in recent years remarkable results in image processing. Many models with different usage of dimensional information were proposed for various datasets. The U-Net [1] structure was developed with 2D convolutions for 2D medical image segmentation and later extended for 3D data. Other approaches included hybrid models with 2D and 3D convolutions together [2].

This work presents a comparison of different techniques and models of CNNs and their results on one dataset, for the task of left ventricle segmentation.

Methods: The dataset consisted of 40 MRI images from 19 different patients, depicting the heart in short-axis. To increase the size of the dataset, an online augmentation approach was chosen, that flipped, rotated, scaled and translated the images.

For a base structure, the U-Net structure was used. Five different models were created, to be used on the available dataset.

The models developed included a 2D U-Net, 3D U-Net, two different hybrid models with 2D and 3D convolutions merged together, and a U-Net consisting of (2+1)D convolutions.

Additionally the effects of parallelization on the training time and possible effects on the accuracy of the results were tested. Lastly the results of training the models segmenting the epicardium and endocardium separately (single-class) or together (multi-class) were compared between the models.

Results: The results showed that parallelization has negative effects on the learning rate of the models. That was caused by bigger batch sizes and input dependent operations of the models. But decreased the time of learning up to 10%.

There was a difference between single-class and multi-class segmentation, that showed that a single-class segmentation are more suitable for this task. Additionally single-class segmentation can use a model that was already trained on the other class, to further improve the results.

The overall comparisons between the dimensions of the models, shown that, for multi-class segmentation, 3D models achieved better results. Where as the overall best results were achieved by a 2D model using single-class segmentation. The hybrid models, showed worse results, without additional benefits, while the (2+1)D model was not able to learn the task of left ventricle segmentation.

Discussion: This presented work shows that for left ventricle segmentation, the most promising results could be achieved with a 2D model, learning single-class segmentation. For multi-class segmentation a 3D model is to favor. Parallelization reduces the training time but has negative side effects.

Funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) – SA 2530/6-1

The authors declare that they have no competing interests.

The authors declare that an ethics committee vote is not required.