# Preserving spatial and quantitative information in unpaired biomedical image-to-image translation

## TL;DR

The paper introduces STABLE, an unpaired image translation method that preserves spatial and quantitative details in biomedical imaging, enabling accurate multimodal integration without needing paired data.

## Contribution

STABLE introduces a novel unpaired image translation approach that preserves spatial and quantitative information through feature consistency and learnable upsampling.

## Key findings

- STABLE preserves spatial alignment and signal intensities better than existing unpaired methods.
- The method successfully translates calcium imaging data from zebrafish brains and virtual histological staining.
- STABLE models complex, non-monotonic relationships across different biomedical image domains.

## Abstract

Analysis of biological samples often requires integrating diverse imaging modalities to gain a comprehensive understanding. While supervised biomedical image translation methods have shown success in synthesizing images across different modalities, they require paired data, which are often impractical to obtain due to challenges in data alignment and sample preparation. Unpaired methods, while not requiring paired data, struggle to preserve the precise spatial and quantitative information essential for accurate analysis. To address these challenges, we introduce STABLE (spatial and quantitative information preserving biomedical image translation), an unpaired image-to-image translation method that emphasizes the preservation of spatial and quantitative information by enforcing information consistency and employing dynamic, learnable upsampling operators to achieve pixel-level accuracy. We validate STABLE across various biomedical imaging tasks, including translating calcium imaging data from zebrafish brains and virtual histological staining, demonstrating its superior ability to preserve spatial details, signal intensities, and accurate alignment compared to existing methods.

•STABLE is an unsupervised image-to-image translation algorithm for biomedical imaging•STABLE preserves spatial and quantitative information from the original image•STABLE can model complex, non-monotonic relationships across different image domains•STABLE outperforms previous methods in a range of biomedical translation tasks

STABLE is an unsupervised image-to-image translation algorithm for biomedical imaging

STABLE preserves spatial and quantitative information from the original image

STABLE can model complex, non-monotonic relationships across different image domains

STABLE outperforms previous methods in a range of biomedical translation tasks

The ability to virtually translate between biomedical imaging modalities holds substantial potential for streamlining biomedical research and diagnostics by offering diverse representations of the same sample. However, supervised translation approaches require paired images, which are often impractical to obtain due to tissue degradation, sample alterations, or incompatible preparation methods. This limitation has led to the development of unpaired image-to-image translation methods. Yet, existing unpaired methods struggle to preserve the precise spatial and quantitative information crucial for biomedical applications, often resulting in misaligned or inaccurate translations. To address these challenges, we introduce STABLE (spatial and quantitative information preserving biomedical image translation), an unpaired image-to-image translation method that enables accurate preservation of spatial alignment and quantitative signals, facilitating reliable multimodal integration of biomedical imaging data.

You et al. propose an unpaired image-to-image translation algorithm for biomedical imaging. They demonstrate that the integration of feature-level consistency with learnable feature upsampling enables the accurate preservation of spatial and quantitative information in a range of biomedical tasks and modalities, including complex, non-monotonic transdomain relationships.