Image Reconstruction in a Manifold of Image Patches: Application to Whole-Fetus Ultrasound Imaging

摘要

We propose an image reconstruction framework to combine a large number of overlapping image patches into a fused reconstruction of the object of interest, that is robust to inconsistencies between patches (e.g. motion artefacts) without explicitly modelling them. This is achieved through two mechanisms: first, manifold embedding, where patches are distributed on a manifold with similar patches (where similarity is defined only in the region where they overlap) closer to each other. As a result, inconsistent patches are set far apart in the manifold. Second, fusion, where a sample in the manifold is mapped back to image space, combining features from all patches in the region of the sample. For the manifold embedding mechanism, a new method based on a Convolutional Variational Autoencoder (β-VAE) is proposed, and compared to classical manifold embedding techniques: linear (Multi Dimensional Scaling) and non-linear (Laplacian Eigenmaps). Experiments using synthetic data and on real fetal ultrasound images yield fused images of the whole fetus where, in average, β-VAE outperforms all the other methods in terms of preservation of patch information and overall image quality.