Generating Healthy Aortic Root Geometries From Ultrasound Images of the Individual Pathological Morphology Using Deep Convolutional Autoencoders

Abstract

In valve-sparing aortic root reconstruction surgery, estimating the individual healthy shape of the aortic root as it was before pathological deformation is a challenging task. However, exactly this estimation is necessary to develop personalized aortic root prostheses. To support the surgeon in this decision making, we present a novel approach to reconstruct the healthy shape of an aortic root based on ultrasound images of its pathologically dilated state using representation learning.The idea is to identify a suitable representation of healthy and pathological aortic root shapes using a supervised variational autoencoder. Then, an image of the dilated root can be encoded, manipulated in the latent space, i.e. shifted towards the distribution of healthy valves, and a synthetic image of this resulting shape can be generated using the decoder.We evaluate our method on an ex-vivo porcine data set and provide a proof-of-concept of our method in a qualitative and quantitavie way. Our results indicate the great potential of reducing a complex shape deformation task to a simple and intuitive shifting towards a specific class. Hence, our method could play an important role in the shaping of personalized implants and is, due to its data-driven nature, not limited to cardiovascular applications but also for other organs.