A Convolutional Autoencoder Approach to Learn Volumetric Shape Representations for Brain Structures

TL;DR

The proposed representation outperforms a state-of-the-art benchmark for brain structures extracted from MRI scans and is invariant to affine transformations, including shifts, rotations and scaling.

Abstract

We propose a novel machine learning strategy for studying neuroanatomical shape variation. Our model works with volumetric binary segmentation images, and requires no pre-processing such as the extraction of surface points or a mesh. The learned shape descriptor is invariant to affine transformations, including shifts, rotations and scaling. Thanks to the adopted autoencoder framework, inter-subject differences are automatically enhanced in the learned representation, while intra-subject variances are minimized. Our experimental results on a shape retrieval task showed that the proposed representation outperforms a state-of-the-art benchmark for brain structures extracted from MRI scans.

Figures (2)

• Figure 1: Proposed architecture. The network consists of a spatial transformer network (STN) and a convolutional autoendoer (CAE). The STN takes input $x_{in}$, a binary segmentation volume, and computes a set of affine transformation parameters $\theta$, which are used to align to the learned reference template $x_{ref}$ using the affine transformation $\mathcal{T}$. The template-aligned scan $x$ is passed through a CAE in order to obtain a shape descriptor $z$ from its bottleneck. The CAE has several residual blocks, where "+input" in the legend indicates a skip connection. Conv:for a $3\times3$ convolution, IN: Instance normalization, LReLU: Leaky Rectified Linear Unit, T.Conv: Transposed convolution, Strd2 Conv: convolution with stride 2. The number of channels is indicated above each layer.
• Figure 2: Lateral and medial views of the learned templates