CRSS: Consistency-regularized semi-supervised spectral unmixing for multicolor fluorescence microscopy images

Multicolor fluorescence microscopy is a cornerstone of biomedical research, enabling simultaneous visualization of multiple subcellular targets. However, inherent spectral overlap between fluorophores causes pervasive detection-channel crosstalk, which distorts quantitative intensity measurements, generates false co-localization signals, and severely limits multiplexing capacity. Traditional unmixing methods rely on pre-calibrated reference spectra and lack adaptability to variable experimental conditions, while deep learning approaches require costly fully labeled fluorescence datasets. We propose CRSS, a semi-supervised blind spectral unmixing method that jointly models spectral-spatial features via the Global-Local Spectral Attention module, adaptively fuses channel-pair features via the Dynamic Feature Fusion module, and imposes bidirectional consistency constraints on unlabeled data using multiple Gaussian perturbations. The method leverages limited labeled and abundant unlabeled data through dual-path synergistic training combining supervised loss, unsupervised reconstruction loss, and multi-perturbation consistency loss. Trained on 344 U-2 OS images with only 9.7 % labeled data (33 samples), CRSS achieves SSIM = 0.972 (just 1.06 % below full supervision) and PSNR = 36.43 dB. It reduces MSE by 23.8 % on cross-domain mouse brain data and generalizes well to BPAE cells. The experimental results show that CRSS achieves a significant improvement in data utilization efficiency through the consistency regularization strategy, and achieves unmixing accuracy close to that of fully supervised methods with only 9.7 % labeled data. This work addresses the high annotation cost challenge for fluorescence spectral data, eliminates the need for laborious spectral calibration, and provides a robust, easy-to-deploy unmixing solution for routine multicolor imaging in cell biology and neuroscience research.