Beer-Lambert Autoencoder for Unsupervised Stain Representation Learning and Deconvolution in Multi-immunohistochemical Brightfield Histology Images
This work addresses stain deconvolution for quantitative analysis in histopathology, particularly for multiplex IHC with more than three stains, offering a data-driven solution to a domain-specific bottleneck.
The paper tackled the problem of separating individual chromogenic stains in multiplex immunohistochemistry (mIHC) RGB histology images, which is under-determined for more than three stains, by introducing an unsupervised autoencoder that learns stain characteristics and produces well-separated concentration maps, resulting in excellent RGB reconstruction and significantly reduced inter-channel bleed-through compared to classical methods.
Separating the contributions of individual chromogenic stains in RGB histology whole slide images (WSIs) is essential for stain normalization, quantitative assessment of marker expression, and cell-level readouts in immunohistochemistry (IHC). Classical Beer-Lambert (BL) color deconvolution is well-established for two- or three-stain settings, but becomes under-determined and unstable for multiplex IHC (mIHC) with K>3 chromogens. We present a simple, data-driven encoder-decoder architecture that learns cohort-specific stain characteristics for mIHC RGB WSIs and yields crisp, well-separated per-stain concentration maps. The encoder is a compact U-Net that predicts K nonnegative concentration channels; the decoder is a differentiable BL forward model with a learnable stain matrix initialized from typical chromogen hues. Training is unsupervised with a perceptual reconstruction objective augmented by loss terms that discourage unnecessary stain mixing. On a colorectal mIHC panel comprising 5 stains (H, CDX2, MUC2, MUC5, CD8) we show excellent RGB reconstruction, and significantly reduced inter-channel bleed-through compared with matrix-based deconvolution. Code and model are available at https://github.com/measty/StainQuant.git.