基于扩散模型的色卡在组织学图像批次校正中的实现

Introduction: Inconsistent tissue preparation, staining, and scanning introduce non-biological variations (batch effects) in histological images. These variations cause machine learning models to learn spurious, site-specific features instead of true pathological patterns, leading to inflated internal performance and poor generalizability and hindering clinical adoption. Current stain normalization methods, such as statistical (Reinhard), color deconvolution (Macenko, Vahadane), and generative adversarial networks, rely on relative normalization to an arbitrary "gold standard," a fundamental limitation. Simple methods often distort important morphological details, while advanced techniques still fail to remove all site-specific signatures. Methods: We introduce a novel generative AI tool for histological image batch correction using a context-aware Stable Diffusion inpainting model to generate a dynamic localized color checker directly on whole slide images (WSIs). The model was trained to inpaint digital representations of a physical color checker (i.e., NIST Traceable Color Transmission Calibration Slide) onto the masked WSI region, preserving the WSI's original color space and establishing a context-aware standard without external references. Trained using 117 Huron and 119 Polaris scanner IHC and H&E WSIs with ground truth scans using the physical NIST color checker, the model enabled subsequent color extraction and image correction to a desired color space. Validation was performed using 78 Huron and 79 Polaris holdout WSIs. The mean-pooled embeddings derived from the Hibou pathology foundational model were used to predict the WSI's original scanner. Results: The proposed method mitigated batch effects by transforming the task from simple normalization into a sophisticated form of image restoration. The model uses its powerful learned prior to reconstruct the image as it should appear under ideal, standardized conditions, mitigating site-specific batching.Prior to correction, this classifier achieved an AUC-ROC of 0.99, indicating the presence of strong, non-biological scanner-specific patterns (batch effects). After normalization, the performance of the same classifier dropped (AUC-ROC = 0.53), confirming the elimination of scanner-specific artifacts. Conclusion: The AI image analysis model offers a robust solution to batch effect challenges in computational pathology, eliminating the need for a reference image during IHC and H&E WSI interpretation. Moving beyond relative color matching, this approach delivers color standardization. By improving robustness to technical batch effects and ensuring AI pathology models are trained on true biological signals, this approach is poised to accelerate the deployment of pathology foundational models across preclinical, translational, and clinical trial settings.