Title: PHASE: Physiology-Aware Hyperspectral Reconstruction via Object-to-Human Domain Adaptation

Abstract:

While hyperspectral imaging provides exceptional, non-invasive insights into human physiology, its widespread clinical adoption is hindered by cumbersome equipment, sluggish data capture speeds, and stringent regulatory requirements. A viable alternative involves deriving hyperspectral data from commonly available RGB or CASSI measurements. However, current methodologies, originally tailored for object-focused imagery, depend on aligning features based on reflectance. These approaches operate under the premise that spectral resemblance equates to semantic consistency—a notion that fails in physiological contexts, where identical RGB appearances can stem from vastly different, intertwined physiological conditions.

This discrepancy necessitates a transition from reflectance-based alignment to representation learning that accounts for physiological factors, anchored in the fundamental principles of light-matter interaction. This transition, however, brings forth two primary obstacles: cross-channel semantic shifts (C1) and the irreversible loss of information inherent in RGB acquisition (C2). To address these issues, we introduce PHASE, a novel framework for hyperspectral reconstruction that reimagines the transfer from object domains to human subjects. PHASE achieves this by separating cross-channel physiological semantics through Physiological Channel Reinterpretation and ensuring that reconstructions remain within physiologically feasible bounds via Physiologically Constrained Alignment.

Evaluated under two distinct source-to-target transfer protocols, PHASE surpasses existing state-of-the-art techniques, achieving improvements of up to +2.20 in SSIM and a reduction of 3.06 in SAM, all while requiring only 1.5% labeled supervision.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC