Guided Filter Regularization for Improved Disentanglement of Shape and Appearance in Diffeomorphic Autoencoders

Abstract

Diffeomorphic and deforming autoencoders have been recently explored in the field of medical imaging for appearance and shape disentanglement. Both models are based on the deformable template paradigm, however they show different weaknesses for the representation of medical images. Diffeomorphic autoencoders only consider spatial deformations, whereas deforming autoencoders also regard changes in the appearance, however no uniform template is generated for the whole training dataset, and the appearance is modeled depending on a very few parameters. In this work, we propose a method that represents images based on a global template, where next to the spatial displacement, the appearance is modeled as the pixel-wise intensity difference to the unified template. To however ensure that the generated appearance offsets adhere to the template shape, a guided filter smoothing of the appearance map is integrated into an end-to-end training process. This regularization significantly improves the disentanglement of shape and appearance and thus enables multi-modal image modeling. Furthermore, the generated templates are crisper and the registration accuracy improves. Our experiments also show applications of the proposed approach in the field of automatic population analysis.