DeepSim: Semantic similarity metrics for learned image registration

TL;DR

DeepSim introduces a semantic similarity metric for image registration that learns dataset-specific features, resulting in higher accuracy, faster convergence, and robustness to noise across various image modalities.

Contribution

It presents a novel semantic similarity metric that improves registration performance and robustness, outperforming existing methods in accuracy and convergence speed.

Findings

Achieves higher registration accuracy than existing methods

Converges faster during optimization

Provides smoother transformations on noisy, low-quality images

Abstract

We propose a semantic similarity metric for image registration. Existing metrics like euclidean distance or normalized cross-correlation focus on aligning intensity values, giving difficulties with low intensity contrast or noise. Our semantic approach learns dataset-specific features that drive the optimization of a learning-based registration model. Comparing to existing unsupervised and supervised methods across multiple image modalities and applications, we achieve consistently high registration accuracy and faster convergence than state of the art, and the learned invariance to noise gives smoother transformations on low-quality images.