Abstract
Background and aim
Automated image recognition based on machine learning methods was proven to be feasible in several medical imaging applications recently. Beside image classification methods to categorize input images for example into healthy or suspicious, image segmentation allows accurate localization of pathologies and thereby facilitates a wide area of applications. Because of the unique composition of every machine learning problem the applicability of image segmentation methods for detecting coronary pathologies in optical coherence tomography images remains unclear. Furthermore, the prediction accuracy of deep learning methods usually depends on vast amounts of training data which are often not available for particular medical questions. Therefore special strategies need to be applied to achieve satisfying results with smaller training datasets. We aimed to investigate the applicability of machine learning methods for plaque detection in coronary OCT images, especially considering the challenge of a small training dataset.
Methods
Originating from a dataset of 104 OCT frames containing calcified plaques, we performed image preprocessing using a custom build OCT image processing software to crop the luminal part as well as the areas outside the circular OCT signal to reduce entropy. Furthermore, plaques were identified and marked by an experienced OCT analyst, drawing plaque-enclosing polygonal masks using the same software. We also performed common image augmentation strategies, primarily applying rotation and zoom operations. Subsequently, we split the samples randomly into training, validation and test datasets (80:10:10%). To train the segmentation model, we fed the training and validation samples into an U-Net Convolutional Neuronal Network implementation with domain-specific adaptions using the RMSprop optimizer based on the publicly available PyTorch library.
Results
After 50 training epochs, we could achieve a prediction accuracy of 74.4% with the current configuration measured by the Sørensen–Dice coefficient comparing the similarity of the predicted plaque masks with the ground truth samples (figure 1 illustrates an exemplary comparison between predicted and ground truth plaque masks).
Exemplary projection of a predicted mask
Conclusion
We were able to show that image segmentation based on machine learning strategies is a feasible way for automated plaque detection in coronary OCT imaging even based on small training datasets. Larger training datasets are necessary to raise prediction accuracy.