Quantification and Mitigation of Site-Specific Differences in Digital Pathology Datasets

This paper addresses the challenges in computational pathology for oral cavity dysplasia, focusing on mitigating site-specific imaging variations in multi-site datasets. Variability in image properties such as brightness, contrast, resolution, and staining techniques confounds detection and classification of histological lesions. In this work, we quantify these differences and implement correction mechanisms including stain normalization and data augmentation to harmonize datasets from multiple international sites. Using principal component analysis, we visualize site-specific differences and show the effect of our standardization approaches. We also train a classifier (Decision Tree) to recognize sites based on image features and show that the ability of the classifier to correctly identify the source site of data is greatly reduced by the standardization approaches, with the most aggressive approach yielding the lowest accuracy (0.70) and F1 scores for each site. These results indicate that both stain normalization and augmentation are necessary to mitigate site-specific dataset differences, and the importance of analysing new site data for its compatibility with existing AI models.