Beer-Lambert Autoencoder for Unsupervised Stain Representation Learning and Deconvolution in Multi-immunohistochemical Brightfield Histology Images
By: Mark Eastwood , Thomas McKee , Zedong Hu and more
Potential Business Impact:
Separates colors in tissue pictures for better analysis.
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.
Similar Papers
Building Trust in Virtual Immunohistochemistry: Automated Assessment of Image Quality
CV and Pattern Recognition
Makes fake stain images as good as real ones.
Cross-Stain Contrastive Learning for Paired Immunohistochemistry and Histopathology Slide Representation Learning
CV and Pattern Recognition
Helps doctors see cancer better with more stain colors.
Cross-Modality Learning for Predicting IHC Biomarkers from H&E-Stained Whole-Slide Images
Image and Video Processing
Predicts cancer protein spots from regular tissue pictures.