BCD-net: Stain separation of histological images using deep variational Bayesian blind color deconvolution
Metadata
Show full item recordAuthor
Yang, Shuowen; Pérez Bueno, Fernando; Castro Macías, Francisco M.; Molina Soriano, Rafael; Katsaggelos, AggelosEditorial
Elsevier
Materia
Deep Bayesian modeling Variational inference Histological images
Date
2024Referencia bibliográfica
Digit. Signal Process. 145 (2024) 104318 [10.1016/j.dsp.2023.104318]
Sponsorship
Project PID2022-140189OB-C22 funded by MCIN / AEI / 10.13039 / 501100011033; Project B-TIC-324- UGR20 funded by FEDER/Junta de Andalucía and Universidad de Granada; Ministerio de Universidades under FPU contract FPU21/01874Abstract
Histological images are often tainted with two or more stains to reveal their underlying structures. Blind Color
Deconvolution (BCD) techniques separate colors (stains) and structural information (concentrations), which is
useful for the processing, data augmentation, and classification of such images.
Classical analytical BCD methods are typically computationally expensive in two distinct ways. First, estimating
the colors and concentrations corresponding to a given image is a time-consuming process. Second, the entire
estimation procedure must be performed independently for each image.
In contrast, Deep Learning (DL) methods involve high training costs, but once trained, they are able to directly
process unseen images. The application of DL to BCD has been limited by the absence of extensive databases
containing ground truth color and concentrations. In this work, we propose BCD-Net, a deep variational
Bayesian neural network for stain separation and concentration estimation. Under this framework, we address
the challenge of lacking ground truth data by leveraging Bayesian modeling and inference techniques.
We propose to use a prior distribution on the stain colors, and a simple flat prior on the concentrations. BCDNet
is trained by maximizing the evidence lower bound of the observed images. The loss function comprises two
essential components: fidelity to the observed images and the Kullback-Leibler divergence between the estimated
posterior distribution of colors and the selected prior.
The model is trained, validated, and tested on two multicenter databases: Camelyon-17 and Warwick stain
separation benchmark. The proposed approach is tested on image reconstruction, stain separation, and cancer
classification. It performs well when contrasted with classical non-amortized methods and offers a substantial
computational time advantage. This marks a significant step forward in the application of DL techniques to
address BCD and paves the way for new approaches.