Medical Hyperspectral Imaging

• Medical hyperspectral imaging is an advanced modality that captures high-resolution spatial and spectral data to analyze tissue biochemistry and morphology.
• It facilitates intraoperative guidance and digital pathology by enabling sub-visual discrimination of tissue types, function, and pathology.
• Integration of machine learning and precise calibration methods offers robust, real-time diagnostics and improved clinical outcomes.

Medical hyperspectral imaging (HSI) is an advanced imaging modality that combines high-resolution spatial and spectral data acquisition to deliver quantitative, pixel-level information on tissue biochemical and morphological properties. By capturing the full reflectance, absorbance, or fluorescence spectrum at each spatial pixel—typically across tens to hundreds of contiguous bands spanning the visible and near-infrared—medical HSI enables “sub-visual” discrimination of tissue classes, function, and pathology. The approach has rapidly evolved from lab-based tissue studies to real-time, intraoperative guidance platforms for surgery, critical care, and digital pathology.

1. Fundamental Principles and System Architectures

At its core, medical HSI leverages tissue-specific light–matter interactions, where endogenous chromophores (e.g., oxygenated/deoxygenated hemoglobin, water, lipids, melanin) exhibit distinct absorption features and scattering profiles as a function of wavelength. The detected signal, parameterized as $$I(\lambda) = I_0(\lambda)\exp[-\mu_a(\lambda)d]T_s(\lambda) + F(\lambda) + N(\lambda)$$, integrates effects of source irradiance $I_0$, absorption coefficient $\mu_a$, optical path $d$, scattering losses $T_s$, fluorescence $F$, and noise $N$ (Hong et al., 11 Aug 2025). In clinical settings, biomedical HSI systems typically operate in the 400–1000 nm range, balancing depth penetration, molecular contrast, and eye safety.

HSI system architectures for medicine include:

• Pushbroom (line-scan) systems: High spectral fidelity (resolving power $R$), used for ex vivo microscopy or static tissue slices.
• Tunable-filter (LCTF/AOTF) systems: Frame-based, flexible spectral selection; limited by acquisition time.
• Snapshot HSI (coded aperture, microlens, or mosaic): Fast, full-field cube acquisition ($\Delta\lambda \approx 10$–$20$ nm), real-time and intraoperative workflows (Wisotzky et al., 2024, Kray et al., 15 Apr 2025).

Calibration is essential: dark/white reference, wavelength alignment (to $\pm0.5$ nm), and geometric correction across bands (Hong et al., 11 Aug 2025, Baumann et al., 2024).

2. Data Processing: Spectral Analysis and Machine Learning

The inherently high-dimensional nature of HSI (spatial size $H\times W$, spectral channels $L$) necessitates dimensionality reduction and advanced modeling. Classical approaches include:

• PCA/MNF: Extract leading spectral features or denoise (noise-aware MNF) (Ezhov et al., 2023, Hong et al., 11 Aug 2025).
• Supervised pixel-wise classification (SVM, Random Forest): Applied on spectral signatures or dimensionality-reduced features (Garcia-Peraza-Herrera et al., 2023).
• Spectral unmixing: Linear model $x = M a + n$ or physically driven unmixing (Monte-Carlo pseudo-endmembers), decomposing pixel spectra into fractions of pure chromophores (oxy-/deoxyhemoglobin, lipids, cytochromes), yielding abundance maps for molecular biomarker localization (Hartenberger et al., 28 Feb 2025).

Deep learning architectures have rapidly advanced:

• 3D CNNs: Joint spectral–spatial convolutions, critical for tissue differentiation (e.g., six-layer, $31\times31\times41$ input, as in parotidectomy tissue mapping) (Wisotzky et al., 2024, Gu et al., 11 Jan 2026).
• U-Net variants: Encoder–decoder for image-based segmentation; can be adapted for both HSI cubes and pseudo-RGB images (Garcia-Peraza-Herrera et al., 2023, Wang et al., 2024).
• Hybrid and attention-based models: Joint 2D/3D, spectral-attention for improved specificity at low label counts (Roddan et al., 31 Jul 2025).

Losses are typically multi-class cross-entropy with $\ell_2$ weight decay; Dice or IoU are used for segmentation evaluation.

3. Interactive and Data-Efficient Segmentation Strategies

Limited annotated data and inter-/intra-patient variability pose persistent challenges, leading to the development of interactive and semi-supervised methods:

• Deep-feature geodesic segmentation: Sparse user scribbles are propagated via geodesic distance in learned feature space (U-Net backbone, $H\times W\times B\rightarrow H\times W\times C$), strongly outperforming RGB or raw-HSI geodesics (Dice up to $0.94$ single-frame, mean $0.842$ over dataset) (Wang et al., 2024).
• Scribble-based and foundation model–guided approaches: SAMSA fuses segmentation confidence from pretrained RGB foundation models with HSI spectral angle similarity (cosine, histogram-equalized), enabling robust, few-shot or zero-shot segmentation with limited annotation—Macro Dice $81.0/93.4\%$ (1/5 clicks, neurosurgical), outperforming standard U-Net or pixel-CNNs (Roddan et al., 31 Jul 2025).
• Semi-supervised label propagation: On intraoperative digital pathology, label-propagation SSL with multiscale spectral-spatial features (MPRI, TensorCSA) permits high accuracy ($\mathrm{BACC}=0.9313$, F1=$0.9235$ with only $1\%$ labeling) and surpasses deep nets with $63\%$ supervision (Kopriva et al., 2024).

4. Clinical and Translational Applications

Medical HSI is deployed in cancer surgery, wound care, perfusion assessment, critical care, and digital pathology:

• Intraoperative tissue differentiation: During parotidectomy, stereo-HSI (41 bands, 400–1000 nm) and 3D CNNs yield $98.7\%$ validation accuracy ($83.4\%$ leave-patient-out), with high nerve/gland sensitivity and vein–muscle confusability (Wisotzky et al., 2024). For low-grade glioma, deep-learning-optimized HSI achieves $89\%$ accuracy with 12 bands, enabling video-rate sensor design (Giannantonio et al., 2023).
• Perfusion and oxygenation mapping: Snapshot light-field HSI combined with rapid SO$_2$ correlation (spectral angle mapping) achieves continuous boundary overlays in $<170$ ms, aligning with surgical judgment (Kray et al., 15 Apr 2025). Quantitative physiological parameter estimation post deep learning–based recalibration matches manual calibration (MAE $3$–$4\%$ vs $5\%$) despite dynamic lighting (Baumann et al., 2024).
• Sepsis and critical care: Palm/finger bedside HSI (TIVITA 2.0, 500–1000 nm, 100 bands) achieves sepsis AUROC $0.80$ and mortality AUROC $0.72$ (HSI only), improving to $0.94$ sepsis AUROC when clinical data are fused (Seidlitz et al., 2024).
• Computational pathology: H&E-stained resection margins; HSI enables $+5$–$6\%$ BACC improvement over RGB, reduces false negatives, and delivers intraoperative margin guidance (Kopriva et al., 2024).

5. System Design, Calibration, and Practical Considerations

Acquisition system design is application-specific:

• Band selection and hardware co-design: Camera selection frameworks employ Monte-Carlo light transport and virtual camera simulation to optimize filter sets and SNR for physiological endpoints (e.g., 4-band oxygenation cameras achieve near-optimal error, $<$10\%) (1904.02709).
• Calibration and workflow integration: Accurate dark/white reference, custom neural recalibrators, and domain-specific stray-light augmentation are key to real-time clinical robustness (Baumann et al., 2024).
• Hybrid modalities: Probe-based systems integrate HSI with structured-light 3D reconstruction, supporting geometry and function overlays in video-rate endoscopy (Lin et al., 2016).
• Snapshot and light-field architectures: Real-time intraoperative use is driven by direct-on-sensor filter mosaics (≥12 bands, 650–780 nm for tumor mapping) (Giannantonio et al., 2023, Kray et al., 15 Apr 2025).

6. Computational Challenges, Robustness, and Emerging Trends

High-dimensionality, cross-device variability, and adversarial fragility are active research areas:

• Spectral–spatial adversarial vulnerability: Attacks exploiting local pixel dependencies and multiscale hierarchies can reduce lesion-specific classification from $82\%\rightarrow8\%$ with minimal visible perturbations; robust training must incorporate structure- and scale-aware regularization (Gu et al., 11 Jan 2026).
• Domain adaptation in HSI reconstruction: SpectralAdapt (SSDA) recovers full HSI cubes from RGB by aligning spectral endmembers and enforcing spectral density–based masking. With $1.5\%$ labeled target images, it achieves SSIM $91.55\%$, SAM $17.76^\circ$, PSNR $30.72$ dB, and downstream segmentation on reconstructed HSI nearly matches ground truth (Wen et al., 17 Nov 2025).

Ongoing directions include computational hyperspectral imaging (joint optics–algorithm design), integration with multi-modal agents (e.g., OCT, MRI), miniaturized probes for endoscopy, and self-supervised pretraining on spectral–spatial data, advancing toward foundation models for cross-device and cross-patient generalization (Hong et al., 11 Aug 2025).

7. Data Standards, Reproducibility, and Clinical Translation

Medical HSI research increasingly mandates:

• Comprehensive metadata: Illumination spectra, acquisition protocols, patient demographics, calibration logs, and annotation protocols with kappa statistics ($\kappa>0.8$ for inter-rater reliability) (Hong et al., 11 Aug 2025).
• Open datasets/code: Standardized repositories (e.g., SPECCHIO, HyperBlood, ColonCancerHSI) facilitate reproducibility and independent benchmarking (Kopriva et al., 2024).
• Safety and clinical integration: Regulatory standards on illumination irradiance ($\leq 20$ mW/cm$^2$), real-time temperature monitoring, and IRB approval underpin intraoperative deployment (Hong et al., 11 Aug 2025).
• Workflow fit: Real-time (<1 s) analysis, minimal annotation burden, hardware/software co-design, and resilience to variable lighting are prerequisites for routine clinical adoption (Baumann et al., 2024, Giannantonio et al., 2023, Kray et al., 15 Apr 2025).

In sum, medical HSI is consolidating as a quantitative, multidimensional tool for label-free diagnostics, surgical guidance, and digital pathology, with deep learning and physical modeling driving clinical maturity. The field is now defined by a convergence of hardware advances, algorithmic innovation, and translational validation (Hong et al., 11 Aug 2025, Wisotzky et al., 2024, Hartenberger et al., 28 Feb 2025, Garcia-Peraza-Herrera et al., 2023).