- Abstract: Longitudinal imaging allows to capture both, the static anatomical structures and the dynamic changes of the morphology due to aging or disease progression. However, common supervised or unsupervised methods for medical imaging do not consider dynamic aspects and process longitudinal data as individual data points. For natural images, algorithms already exist that learn a representation from videos. In retinal imaging, however, the temporal sampling resulting from follow-up to disease progression is much lower than in videos. Predictions are therefore more ambiguous and prone to noise. We propose a deep learning approach to overcome these challenges, which allows us to understand the underlying morphological organization and its changes over time, and to discover abnormalities and pathologic evolutions. Our data-driven approach learns a feature representation from unlabeled longitudinal images by predicting the unobserved subsequent image within a series of observations. Several sources of noise, such as imaging noise, misalignment of follow-up images or motion artifacts aggravates the direct prediction of the target image. Thus, we propose to adapt a Conditional Variational Autoencoder (CVAE) [Kingma and Welling] to learn representative static and dynamic features that are robust to noise and uncertainty.
- Keywords: Representation learning, Longitudinal, Autoencoder, Variational Autoencoder, Retinal Imaging, Dynamic Features, Deep Learning
- Author affiliation: Christian Doppler Laboratory for Ophthalmic Image Analysis, Department of Ophthalmology and Optometry, Medical University of Vienna