Learning to Predict the Optimal Template in Stain Normalization for Histology Image Analysis

Histology images are the gold standard for tumor diagnosis, and numerous studies have investigated deep learning methods for their analysis. However, stain heterogeneity, resulting from variations in staining protocols and batches across laboratories, impairs the generalization ability of these methods. Although stain normalization can mitigate this issue, current approaches rely on the manual selection of template images, which is time-consuming, requires domain expertise, and is susceptible to subjective biases. To address this challenge, we propose Learnable Automatic Normalization (LAN), an innovative two-stage method that automates the stain normalization process. In the first stage, a deep network in LAN learns to predict the optimal stain normalization template for each histology image, eliminating the need for manual template selection. Then, LAN generates virtually normalized histology samples using the predicted template. These normalized samples are then used in the second stage for a downstream task, with the supervision signal from this stage updating the network in the first stage. We evaluated LAN on the NCT-CRC-HE-100K-NORMAL dataset and demonstrated its superior performance in terms of accuracy compared to existing stain normalization methods. By automating and optimizing the stain normalization process, LAN has the potential to streamline the workflow in histology image analysis, ultimately contributing to more efficient and accurate pathological diagnoses. The code is available at https://github.com/Christyshiling/LAN.