Rapid tissue oxygenation mapping from snapshot structured-light images with adversarial deep learning

Mason T. Chen, Nicholas J. DurrarXiv 2020 · arXiv

Abstract

Tissue oxygenation ($StO_2$) is a critical biomarker for assessing tissue viability in conditions such as sepsis, diabetes, and surgical recovery. While Spatial Frequency Domain Imaging (SFDI) provides a non-contact method for mapping $StO_2$, conventional multi-image acquisition is slow and motion-sensitive. This work introduces OxyGAN, an end-to-end deep learning framework using conditional generative adversarial networks (cGANs) to map $StO_2$ from a single snapshot of structured light. OxyGAN achieves over 96% accuracy and enables 25Hz video-rate imaging, outperforming traditional physics-based and hybrid methods by approximately 25%.

Introduction

Mapping tissue oxygenation ($StO_2$) over a wide field of view is essential for real-time clinical assessment. Spatial Frequency Domain Imaging (SFDI) is a leading technique for this purpose, but its reliance on multiple sequential images creates a bottleneck, making it susceptible to patient movement and unsuitable for high-speed applications. Single-snapshot methods (SSOP) attempt to solve this but often suffer from image artifacts and reduced accuracy.

Comparison of traditional SFDI, hybrid GANPOP, and the proposed end-to-end OxyGAN workflows.

Methods

OxyGAN utilizes a conditional Generative Adversarial Network (cGAN) to directly map raw structured-light images to $StO_2$ maps. The generator is a fusion network combining U-Net's spatial context capabilities with ResNet's residual learning efficiency, allowing the model to learn both local features and high-level spatial structures.

The OxyGAN architecture featuring a hybrid U-Net/ResNet generator and a patch-based discriminator.

The model processes a three-channel input: structured-light images at 659 nm and 851 nm, and a reference "checkerboard" channel to account for hardware system drift. By employing an end-to-end learning strategy, OxyGAN avoids the compounding errors typical of multi-step "hybrid" models that first calculate intermediate optical properties like absorption coefficients. To achieve real-time performance, the model was optimized using NVIDIA TensorRT, reaching an inference latency of approximately 0.04s.

Results and Discussion

The model demonstrated high accuracy (96.5%) on human feet and maintained strong performance (93.0%) on unseen tissue types, such as pig colon and human hands across various skin types. OxyGAN significantly reduced "fringe" artifacts common in single-snapshot imaging by learning the underlying distribution of tissue structures rather than relying on simple frequency filtering.

Compared to traditional SSOP and hybrid deep learning models, OxyGAN provided a performance gain of approximately 25%. The ability to produce clean, accurate oxygenation maps at 25Hz enables its use in dynamic environments like the operating room for real-time surgical guidance.

Conclusion

OxyGAN provides a robust, high-speed solution for tissue oxygenation mapping. By integrating adversarial learning with a simplified snapshot acquisition protocol, it bridges the gap between high-fidelity multi-image SFDI and the speed requirements of clinical practice. Future work will explore more diverse datasets and alternative illumination strategies to further enhance the system's accessibility and robustness.

Summary updated

Abstract

Spatial frequency domain imaging (SFDI) is a powerful technique for mapping tissue oxygen saturation over a wide field of view. However, current SFDI methods either require a sequence of several images with different illumination patterns or, in the case of single snapshot optical properties (SSOP), introduce artifacts and sacrifice accuracy. To avoid this tradeoff, we introduce OxyGAN: a data-driven, content-aware method to estimate tissue oxygenation directly from single structured light images using end-to-end generative adversarial networks. Conventional SFDI is used to obtain ground truth tissue oxygenation maps for ex vivo human esophagi, in vivo hands and feet, and an in vivo pig colon sample under 659 nm and 851 nm sinusoidal illumination. We benchmark OxyGAN by comparing to SSOP and to a two-step hybrid technique that uses a previously-developed deep learning model to predict optical properties followed by a physical model to calculate tissue oxygenation. When tested on human feet, a cross-validated OxyGAN maps tissue oxygenation with an accuracy of 96.5%. When applied to sample types not included in the training set, such as human hands and pig colon, OxyGAN achieves a 93.0% accuracy, demonstrating robustness to various tissue types. On average, OxyGAN outperforms SSOP and a hybrid model in estimating tissue oxygenation by 24.9% and 24.7%, respectively. Lastly, we optimize OxyGAN inference so that oxygenation maps are computed ~10 times faster than previous work, enabling video-rate, 25Hz imaging. Due to its rapid acquisition and processing speed, OxyGAN has the potential to enable real-time, high-fidelity tissue oxygenation mapping that may be useful for many clinical applications.
BibTeX
@misc{chen2020rapid,
  author = {Mason T. Chen and Nicholas J. Durr},
  title = {{Rapid tissue oxygenation mapping from snapshot structured-light images with adversarial deep learning}},
  year = {2020},
  eprint = {2007.00760},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2007.00760},
}
Figure