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

Near-infrared diffuse reflectance hyperspectral imaging (HSI), as an emerging noninvasive imaging technology, captures organ-specific spectral information, but the spectrum may be different between individuals and changes with physiological states. A triple-perception adaptive network, termed as TriAd, is developed for in vivo organ recognition in zebrafish by correcting interindividual variability arising from variations in experimental conditions and developmental stages. TriAd incorporates three complementary branches designed to extract organ-specific hyperspectral features from images. A graph convolutional network branch is employed for capturing global spectral-structural dependencies, a 2D convolutional neural network branch is utilized for extracting local spatial-spectral features, and a discrete wavelet transform branch is adopted for resolving different-scale spectral components. The maximum mean discrepancy is used across the three perceptual branches to reduce cross-domain differences induced by experimental and developmental variations. Evaluation with a near-infrared diffuse reflectance HSI data set comprising 12 organs from zebrafish at two developmental stages demonstrated the superior classification performance of TriAd, outperforming conventional chemometric calibration transfer methods, adversarial domain adaptation approaches, and deep learning models. TriAd provides a framework for hyperspectral analysis under variations in experimental conditions and biological development with potential applications in living object analysis and noninvasive medical imaging.