Learnable real-time inference of molecular composition from diffuse spectroscopy of brain tissue

Ivan Ezhov, Luca Giannoni, Ivan Iliash, Felix Hsieh, Charly Caredda, Fred Lange, Ilias Tachtsidis, Daniel Rueckert, Lange, Fred, Tachtsidis, Ilias, Rueckert, Daniel

Abstract

Diffuse optical modalities such as broadband near-infrared spectroscopy (bNIRS) and hyperspectral imaging (HSI) represent a promising alternative for low-cost, non-invasive, and fast monitoring of functional and structural properties of living tissue. Particularly, the possibility of extracting the molecular composition of the tissue from the optical spectra in real-time deems the spectroscopy techniques as a unique diagnostic tool. However, no established method exists to streamline the inference of the biochemical composition from the optical spectrum for real-time applications such as surgical monitoring. In this paper, we analyse a machine learning technique for fast and accurate inference of changes in the molecular composition of brain tissue. We reconsider and propose modifications to the existing learnable methodology based on the Beer-Lambert law, which analytically connects the spectra with concentrations. We evaluate the method's applicability to linear and non-linear formulations of the Beer-Lambert law. The approach is tested on real data obtained from the bNIRS- and HSI-based optical monitoring of brain tissue. The results demonstrate that the proposed method enables real-time molecular composition inference while maintaining the accuracy of traditional linear and non-linear optimization solvers. Preliminary findings show that Beer-Lambert law-based spectral unmixing allows to contrast brain anatomy semantics such as the vessel tree and tumor area.

Introduction

In neurosurgery, the precise delineation between tumorous and healthy tissue is critical. Diffuse optical tools, including broadband near-infrared spectroscopy (bNIRS) and hyperspectral imaging (HSI), offer a non-invasive means to assess tissue biochemistry by analyzing light absorption and scattering. However, the "inverse problem"—converting raw light spectra into molecular concentrations (e.g., oxy-hemoglobin, cytochrome-c-oxidase)—traditionally relies on complex physics-based optimization. While accurate, these solvers are too slow for real-time intra-operative use, especially for HSI data which contains millions of pixels. This study proposes a machine learning approach to replace slow optimization with fast neural network inference.

Methods

The proposed methodology utilizes Multi-Layer Perceptrons (MLPs) to learn the mapping between optical spectra and molecular concentration changes, grounded in the modified Beer-Lambert law.

Overview of the proposed pipeline: from dataset generation using the Beer-Lambert law to neural network training and real-time inference.

Model Complexity and Training Strategies

The researchers investigated both linear models (absorption only) and non-linear models (incorporating scattering). To train these models, two distinct strategies were evaluated:

  1. Strategy (a): Training on purely synthetic data generated from physical equations.
  2. Strategy (b): Training on real spectra labeled by a traditional, slow least-squares solver.
Comparison of training strategies: Strategy (a) uses synthetic data, while Strategy (b) incorporates real data labeled by a traditional solver to better handle experimental noise.

Hyperparameter optimization was performed using AutoML (Ray Tune) to ensure efficient network architectures.

Datasets

Validation was conducted on two datasets:

  • Broadband NIRS: Time-series data from 27 piglets undergoing hypoxia-ischemia, tracking molecular changes over several hours.
  • Hyperspectral Imaging (HSI): Data from the HELICoiD project, featuring intra-operative brain images from 22 human patients with glioma.

Results and Discussion

The neural network approach achieved a massive speedup, processing 100,000 spectra in just 0.4 milliseconds on a mobile GPU. For HSI, the entire pipeline took approximately 3 seconds, a significant improvement over traditional solvers that require minutes or hours.

Key findings include:

  • Importance of Scattering: Including scattering in non-linear models improved spectral fit by 15.7% on average, particularly in the 740nm–780nm range.
  • Strategy Effectiveness: While synthetic training is sufficient for linear models, Strategy (b) (training on solver-labeled real data) proved superior for complex non-linear scenarios by accounting for instrumentation noise.
  • Clinical Utility: The resulting molecular maps provided enhanced contrast for blood vessels and tumor boundaries compared to standard RGB images.

Limitations

Current limitations include the assumption of a constant light pathlength and the lack of a standardized chromophore set for HSI tumor imaging. Future work will focus on hardware scaling to achieve video-rate biochemical imaging.

Conclusion

By transforming a physics-based optimization problem into a learnable pattern-recognition task, this method enables real-time biochemical mapping of brain tissue, bringing augmented-reality surgical guidance closer to clinical reality.

Summary updated

Abstract

Diffuse optical modalities such as broadband near-infrared spectroscopy (bNIRS) and hyperspectral imaging (HSI) represent a promising alternative for low-cost, non-invasive, and fast monitoring of functional and structural properties of living tissue. Particularly, the possibility of extracting the molecular composition of the tissue from the optical spectra in real-time deems the spectroscopy techniques as a unique diagnostic tool. However, no established method exists to streamline the inference of the biochemical composition from the optical spectrum for real-time applications such as surgical monitoring. In this paper, we analyse a machine learning technique for fast and accurate inference of changes in the molecular composition of brain tissue. We reconsider and propose modifications to the existing learnable methodology based on the Beer-Lambert law, which analytically connects the spectra with concentrations. We evaluate the method's applicability to linear and non-linear formulations of the Beer-Lambert law. The approach is tested on real data obtained from the bNIRS- and HSI-based optical monitoring of brain tissue. The results demonstrate that the proposed method enables real-time molecular composition inference while maintaining the accuracy of traditional linear and non-linear optimization solvers. Preliminary findings show that Beer-Lambert law-based spectral unmixing allows to contrast brain anatomy semantics such as the vessel tree and tumor area.
BibTeX
@article{ezhov2024learnable,
  author = {Ivan Ezhov and Luca Giannoni and Ivan Iliash and Felix Hsieh and Charly Caredda and Fred Lange and Ilias Tachtsidis and Daniel Rueckert and Lange, Fred and Tachtsidis, Ilias and Rueckert, Daniel},
  title = {{Learnable real-time inference of molecular composition from diffuse spectroscopy of brain tissue}},
  journal = {Journal of Biomedical Optics},
  year = {2024},
  doi = {10.1117/1.jbo.29.9.093509},
}
Figure