Vis-NIR spectroscopy and deep learning for non-invasive in vivo detection of acute liver injury

Acute liver injury (ALI) is accompanied by coordinated alterations in hepatic perfusion, lipid metabolism, and protein structure, which collectively modulate tissue optical properties. Despite the clinical need for early detection, capturing these changes noninvasively under in vivo conditions remains challenging due to strong light scattering, background interference, and spectral redundancy inherent in multi-layered biological tissues. In this study, we investigate the capability of visible and near-infrared (Vis-NIR) diffuse reflectance spectroscopy (350-2500 nm) to characterize spectral signatures associated with ALI. A mouse model of ALI was established via intraperitoneal injection of 300 mg/kg acetaminophen (APAP), and transcutaneous (non-invasive) abdominal diffuse reflectance spectra were acquired. To address the high-dimensional, low-sample-size (HDLSS) characteristics and limited signal-to-noise ratio of spectral data, standardized preprocessing and an embedded interpretable deep learning framework were employed to enable end-to-end spectral feature selection. Model performance was evaluated using stratified cross-validation, achieving a recall of 0.876 ± 0.085 and a stable F1-score of 0.814 ± 0.037. Repeated feature selection revealed stable wavelength clusters, notably at 1680-1689 nm and 2130-2139 nm, rather than isolated peaks. These clusters align with known vibrational contributions from lipids and proteins, as well as hemodynamic-related optical contrast. Such multi-band signatures reflect the complex physiological changes associated with acute liver injury (ALI). By identifying these biologically plausible markers through the skin, this study demonstrates that Vis-NIR spectroscopy provides a non-invasive tool for ALI detection without requiring one-to-one molecular assignments.