Exploring Deep Registration Latent Spaces

TL;DR

This paper investigates the interpretability of deep learning-based registration methods by decomposing their latent spaces into anatomically meaningful transformations, enhancing understanding of these complex models.

Contribution

It introduces a novel approach using linear projection to decompose registration latent spaces into interpretable, anatomically aware components, demonstrated on lung and hippocampus MRI datasets.

Findings

Latent spaces can be decomposed into orthogonal, anatomically meaningful transformations.

The approach reveals properties of the convoluted latent spaces in registration pipelines.

Decomposition aids in understanding the geometrical transformations captured by deep models.

Abstract

Explainability of deep neural networks is one of the most challenging and interesting problems in the field. In this study, we investigate the topic focusing on the interpretability of deep learning-based registration methods. In particular, with the appropriate model architecture and using a simple linear projection, we decompose the encoding space, generating a new basis, and we empirically show that this basis captures various decomposed anatomically aware geometrical transformations. We perform experiments using two different datasets focusing on lungs and hippocampus MRI. We show that such an approach can decompose the highly convoluted latent spaces of registration pipelines in an orthogonal space with several interesting properties. We hope that this work could shed some light on a better understanding of deep learning-based registration methods.