Identifying chromophore fingerprints of brain tumor tissue on hyperspectral imaging using principal component analysis

Ivan Ezhov, Luca Giannoni, Suprosanna Shit, Frédéric Lange, Florian Kofler, Bjoern Menze, Ilias Tachtsidis, Daniel Rueckert

Abstract

Hyperspectral imaging (HSI) is an optical technique that processes the electromagnetic spectrum at a multitude of monochromatic, adjacent frequency bands. The wide-bandwidth spectral signature of a target object's reflectance allows fingerprinting its physical, biochemical, and physiological properties. HSI has been applied for various applications, such as remote sensing and biological tissue analysis. Recently, HSI was also used to differentiate between healthy and pathological tissue under operative conditions in a surgery room on patients diagnosed with brain tumors. In this article, we perform a statistical analysis of the brain tumor patients' HSI scans from the HELICoiD dataset with the aim of identifying the correlation between reflectance spectra and absorption spectra of tissue chromophores. By using the principal component analysis (PCA), we determine the most relevant spectral features for intra- and inter-tissue class differentiation. Furthermore, we demonstrate that such spectral features are correlated with the spectra of cytochrome, i.e., the chromophore highly involved in (hyper) metabolic processes. Identifying such fingerprints of chromophores in reflectance spectra is a key step for automated molecular profiling and, eventually, expert-free biomarker discovery.

Introduction

In the high-stakes environment of neurosurgery, distinguishing between a malignant glioma and healthy brain tissue is notoriously difficult. Hyperspectral imaging (HSI) offers a non-invasive solution by capturing hundreds of narrow wavelength bands, creating a "spectral fingerprint" for every pixel in an image. However, researchers face a significant challenge known as the inverse problem. While tissue reflectance is dictated by the absorption of specific molecules (chromophores like hemoglobin or cytochromes), working backward from a reflection spectrum to a precise chemical composition is often "ill-posed." Because the absorption spectra of different molecules overlap and are not mathematically independent, multiple chemical combinations can produce identical visual results. This study seeks a model-agnostic method to identify which specific molecules drive the differences observed in HSI scans of brain tumors.

The HELICoiD Dataset

The study utilized the HELICoiD dataset, a prominent public benchmark containing HSI scans of patients undergoing surgery for Grade IV glioblastoma. The images include four sparsely labeled classes: tumor tissue, normal healthy tissue, blood vessels, and background. The spectral range covers 826 bands across the visible and near-infrared spectrum. For this analysis, the authors preselected twelve patients and extracted a balanced set of 30,000 spectra across the three primary tissue classes.

Visualization of the HELICoiD dataset showing synthetic RGB images and masks for tumor, normal tissue, and blood vessels.

Methodology

The authors applied Principal Component Analysis (PCA) to the extracted spectra. Instead of forcing the data to fit a pre-defined biological model, they used PCA to find the "principal components" (PCs)—the axes that capture the most variance in the data.

  1. Data Extraction: 30,000 spectra were pulled from HSI cubes representing tumor, normal tissue, and blood vessels.
  2. Dimensionality Reduction: PCA reduced the 826 spectral bands into a few key components.
  3. Correlation Analysis: The weights assigned to different wavelengths in these PCs were compared against known absorption spectra of various chromophores (oxy-hemoglobin, deoxy-hemoglobin, and cytochromes) sourced from established biological repositories.

Key Insights: The Cytochrome Fingerprint

A primary finding of the study was the power of the first principal component (PC1), which alone explained more than 98% of the variance in the dataset. This suggests that a single underlying biological factor is responsible for the vast majority of the spectral differences between healthy and cancerous brain tissue.

Comparison of HSI reflectance spectra for different tissue types against the absorption coefficients of biological chromophores.

The weights of PC1 showed a remarkably close match to the absorption profile of reduced cytochrome-c-oxidase (redCCO), particularly in the 500–600 nm range. Cytochromes are central to mitochondrial energy production; because tumors are hyper-metabolic, they possess abnormal concentrations of these proteins. PCA successfully captured this biological reality without prior instruction to look for it.

Results and Discussion

The analysis confirmed that spectral variance was highest when comparing tumor tissue to blood or normal tissue, reinforcing that the cytochrome signature is a primary differentiator for malignancy. Even within a single tissue class (intra-class variance), the highest variance still aligned with cytochrome absorption, reflecting the known heterogeneity of gliomas where some areas are actively proliferating while others are necrotic.

Statistically, PCA weights for "tumor vs. normal" and "tumor vs. blood" datasets were significantly higher than for "normal vs. blood," further validating the cytochrome signature as a specific marker for the tumor microenvironment.

Conclusion

While PCA is excellent for identifying correlations, the non-bijective nature of the mapping remains a hurdle due to overlapping chromophore peaks. Future work will focus on expanding the study to larger patient cohorts and more diverse chromophores. Ultimately, integrating these PCA-identified "fingerprints" into automated molecular profiling tools could lead to expert-free biomarker discovery, allowing computers to identify tumor margins with high biochemical precision during live surgery.

Summary updated

Abstract

Hyperspectral imaging (HSI) is an optical technique that processes the electromagnetic spectrum at a multitude of monochromatic, adjacent frequency bands. The wide-bandwidth spectral signature of a target object's reflectance allows fingerprinting its physical, biochemical, and physiological properties. HSI has been applied for various applications, such as remote sensing and biological tissue analysis. Recently, HSI was also used to differentiate between healthy and pathological tissue under operative conditions in a surgery room on patients diagnosed with brain tumors. In this article, we perform a statistical analysis of the brain tumor patients' HSI scans from the HELICoiD dataset with the aim of identifying the correlation between reflectance spectra and absorption spectra of tissue chromophores. By using the principal component analysis (PCA), we determine the most relevant spectral features for intra- and inter-tissue class differentiation. Furthermore, we demonstrate that such spectral features are correlated with the spectra of cytochrome, i.e., the chromophore highly involved in (hyper) metabolic processes. Identifying such fingerprints of chromophores in reflectance spectra is a key step for automated molecular profiling and, eventually, expert-free biomarker discovery.
BibTeX
@misc{ezhov2023identifying,
  author = {Ivan Ezhov and Luca Giannoni and Suprosanna Shit and Frédéric Lange and Florian Kofler and Bjoern Menze and Ilias Tachtsidis and Daniel Rueckert},
  title = {{Identifying chromophore fingerprints of brain tumor tissue on hyperspectral imaging using principal component analysis}},
  year = {2023},
  eprint = {2301.05233},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2301.05233},
}
Figure