Abstract
Hyperspectral Raman imaging (HRI) enables label-free molecular mapping of cells but is constrained by weak Raman scattering signals that require long integration times. We present a depthwise separable three-dimensional MultiResU-Net that performs joint spatial-spectral denoising on full HRI data cubes. By integrating multi-scale feature fusion and depthwise separable convolutions, the network effectively captures spatial-spectral correlations, achieving accurate reconstruction of both spectral features and cellular morphology. Validated on synthetic and experimentally acquired HRI datasets, the proposed method substantially enhances signal quality while preserving structural fidelity. It enables rapid cellular Raman imaging at short integration times without sacrificing spectral integrity, outperforming traditional filters and one-dimensional neural networks. This approach provides an efficient and generalizable computational framework for accelerating HRI, supporting broader applications in high-throughput and dynamic biomedical analysis.