Identification of electrical rotational activity in noisy cardiac tissue recordings using a deep neural network

Abstract

Type of funding sources: None. Deep learning is increasingly used in modern biomedical research and applications due to the substantial availability of large clinical datasets. These approaches are invaluable in tasks involving noisy imaging data, such as tumour segmentation in histological images. In cardiology, a deep learning approach could be helpful in real-time tracking of the sources of arrhythmia, i.e. electrical rotational activity in the heart. However, the existing optical or electrophysiological recordings that could be used for training such a model are recorded under highly variable conditions and are not always annotated, thereby requiring data augmentation. To use deep neural networks trained on synthetic data to obtain concise (low dimensional) representations of noisy optical mapping recordings of cardiac arrhythmias and rapidly locate spiral wave centres. To overcome the lack of experimental training data, a digital twin of a neonatal rat ventricular cardiomyocyte monolayer was used to create a large synthetic training dataset of noiseless spiral wave recordings. Spiral wave centres were detected and labelled by making use of classical algorithms which are proven to work well on noiseless data. After labelling the centres, noise was added to the spiral wave recordings to simulate realistic experimental measurements. Subsequently, these data were fed into three different deep learning architectures: 1) a variational auto-encoder (VAE) to denoise optical mapping recordings of cardiac arrhythmias in an unsupervised manner, 2) a convolutional neural network (CNN) to detect the spiral centres, and 3) a combination of both to denoise the recording and detect centres simultaneously. After training on synthetic datasets, each architecture could accurately predict what it was designed for (noiseless wave fronts, spiral centres including chirality, or both) on both synthetic and experimental data. These spiral centre detection results were compared with 5 classical methods of denoising and spiral centre detection for accuracy and speed. Our method was as accurate as the best performing yet slow classical algorithm, which can only detect centres after observing a full rotation cycle (~300ms). At the same time, it was as fast as the fastest classical method, needing only 30ms after enabling the algorithm to detect spiral wave centres. This allows quasi-real-time tracking of arrhythmic sources. This study reveals that modern deep learning strategies in combination with synthetic simulation datasets can be used on experimental measurements to aid in the development of new technologies, here applied to the detection of spiral wave centres in optical mapping recordings of cardiac arrhythmias. Given the combination of speed and accuracy at which these algorithms produce results, further exploration and refinement may improve the identification of targets for catheter ablation, thereby potentially improving the outcome.