A semi-supervised autoencoder framework for joint generation and classification of breathing

One of the main problems with biomedical signals is limited amount patient-specific data and significant time needed to record sufficient number samples for diagnostic treatment purposes. In this study, we present a framework simultaneously generate classify series based on modified Adversarial Autoencoder (AAE) algorithm one-dimensional convolutions. Our work breathing series, specific motivation capture motion during radiotherapy lung cancer treatments. First, explore potential in using Variational (VAE) AAE algorithms model from individual patients. We then extend allow joint semi-supervised classification generation different types within single framework. To simplify modeling task, introduce pre-processing post-processing compressing that transforms multi-dimensional into vectors containing position values, which are transformed back through an additional neural network. The resulting models able realistic varied breathing. By incorporating 4% 12% labeled training, our outperforms other purely discriminative networks classifying baseline shift irregularities dataset completely training set, achieving average macro F1-score 94.91% 96.54%, respectively. knowledge, presented first approach unifies type data, enabling both computer aided diagnosis augmentation