StainOptimizer: A Whole-Process Evaluation Approach with Interpretability Framework for Optimizing Virtual Staining Methods

Virtual staining shows great potential in computational histopathology but faces clinical challenges due to the limitations in output quality (e.g., unclear tissue details and inconsistent staining styles). While current studies have developed various virtual staining methods, most quality evaluations of these methods are based on the similarity comparisons of generated images. This single-metric evaluation fails to elucidate the root causes of the limitations, leading to untargeted model optimization and unreliable clinical deployment. Therefore, it is essential to conduct a comprehensive evaluation of virtual staining. Here, this article introduces a whole-process evaluation approach for optimizing virtual staining methods, called StainOptimizer. It employs a three-stage interpretability framework (pre-modeling, modeling, post-modeling) to dissect the contributions of each component, and adapts evaluation metrics to clinical scenarios via dynamic indicator selection. Unlike previous methods, StainOptimizer compiles all evaluation indicators into a comprehensive report for model optimization. To verify the effectiveness of StainOptimizer, this article applies it to compare the performance of two mainstream methods: pix2pix and cycleGAN. The evaluation revealed a visual superiority of cycleGAN; however, it exhibited poorer structural details due to limitations in data quality and generator architecture. Guided by StainOptimizer, the optimized model achieved improvements in tissue detail capture and staining consistency.