DynamicFusionNet: H&E stain normalization using image disentangling

With the rapid advancement of digital pathology technologies, stain normalization has become a critical step in addressing staining variations across histopathological images. Although existing methods have partially resolved the issue of inconsistent staining, challenges such as detail loss, artifact generation, and insufficient structural information extraction remain. These limitations hinder the effectiveness and stability of these methods in practical applications. To address these challenges, this chapter introduces a novel image-disentangling-based stain normalization network, which leverages a selfsupervised learning approach to effectively learn in the target domain and can be applied to stain normalization across unknown domain. Two innovative modules were designed as part of the network. The Multi-scale Feature Enhancement Module (MFEM) combines multi-scale convolution with depthwise separable convolution to accurately extract structural information at both microscopic and macroscopic levels, providing high-quality structural features for subsequent stain normalization. The Dynamic Feature Adjuster (DFA) employs dynamic feature adjustment mechanisms and adaptive sampling techniques to overcome common issues in existing methods, such as artifact generation during the reconstruction of staining and structural features, further improving the accuracy and stability of stain normalization. These innovations enable the network to adapt more effectively to complex and diverse histopathological image data, providing an efficient and precise solution for stain normalization tasks.