A Triple-Perception Adaptive Network for In Vivo Organ Recognition Using Diffuse Reflectance Hyperspectral Imaging

Area of Science:

• Biomedical optics
• Machine learning for biological imaging
• Noninvasive in vivo imaging

Background:

• Near-infrared diffuse reflectance hyperspectral imaging (HSI) captures organ-specific spectral data.
• Interindividual spectral variability arises from physiological states and experimental conditions.
• Accurate organ recognition requires addressing these spectral variations.

Purpose of the Study:

• Develop a novel adaptive network (TriAd) for in vivo organ recognition in zebrafish.
• Correct for interindividual variability in HSI data caused by experimental and developmental differences.
• Enhance the accuracy and robustness of noninvasive organ identification.

Main Methods:

• A triple-perception adaptive network (TriAd) with three branches: graph convolutional network, 2D convolutional neural network, and discrete wavelet transform.
• Utilizing maximum mean discrepancy to minimize cross-domain differences between perceptual branches.
• Training and evaluation on a near-infrared diffuse reflectance HSI dataset of 12 zebrafish organs across two developmental stages.

Main Results:

• TriAd demonstrated superior classification performance for in vivo organ recognition in zebrafish.
• The network effectively corrected for interindividual variability and cross-domain differences.
• TriAd outperformed existing chemometric calibration transfer, adversarial domain adaptation, and deep learning methods.

Conclusions:

• TriAd offers a robust framework for hyperspectral analysis in the presence of experimental and developmental variations.
• The developed network shows significant potential for applications in living object analysis and noninvasive medical imaging.
• This approach advances noninvasive organ recognition through adaptive hyperspectral data processing.