Label-free sub-cellular segmentation via deep learning-based hyperspectral stimulated Raman microscopy

Accurate segmentation of sub-cellular components is crucial for understanding cellular metabolic functions. Hyperspectral Stimulated Raman Scattering (HSRS) microscopy enables label-free molecular chemical imaging of different components by providing spatial and spectral information, thus enabling researchers to understand complex cellular structures at the molecular level. However, traditional spectral Phasor analysis, widely used for sub-cellular segmentation in HSRS images, considers only the spectral dimension while ignoring the spatial distribution of HSRS images. The results heavily rely on manual parameter tuning without high robustness. To address these limitations, we developed a novel deep learning model based on a hybrid-UNet architecture. This model is designed to simultaneously harness both spatial and spectral features effectively from HSRS data using three strategies. First, the encoder incorporates optional Multi-Scale Residual 3D Convolution Modules to extract spectral features, together with preserving spatial details by residual connections. Second, a spectral compression module at the deepest layer reduces computational complexity. Final, the attention-weighted pooling in layer connections emphasizes the spectral features and compresses the spectral vector into a compact embedding that connects the encoder and decoder. This architecture enables tailored feature extraction for variable organelles, such as nucleus and lipid droplets. We evaluated our model on HSRS images of exfoliated cells from gastric cancer. Using an optimized Phasor method as the gold standard, our approach achieved Dice coefficients of 0.9184 and 0.9433 for nucleus and lipid droplets segmentation, respectively. The results demonstrate superior performance over mainstream segmentation models like nnUNet and 3D-UNet, and highlight our proposed modules through comprehensive ablation studies. Our method provides an efficient, robust, and automated tool for sub-cellular segmentation, paving the way for high-throughput analysis in label-free chemical imaging.