Fig. 1: Experimental design and pipeline overview.

The existing Digital Pathology assessment pipeline is illustrated in (a). In this study, we developed an integrated and comprehensive Digital Pathology image analysis pipeline, which is powered by A!MagQC (image quality assessment software), A!HistoClouds (Digital Pathology image viewer, annotation platform, and database), and an Artificial Intelligence model that can detect and grade prostate cancer for images scanned by multiple scanners, as shown in (b). The workflow of the pathologist-AI interaction (PAI) is presented in (c). After a base AI model is trained, it is applied to new data and generate pseudo annotations on Whole Slide Images. Pathologists review and modify the pseudo annotations using A!HistoClouds, and the corrected annotations are fed back to the model for further training. This process is repeated until the model achieves high accuracy. After meeting criteria, the model can generate accurate outcomes and assist pathologists in clinical diagnosis. In AI digital pathology, AI model performance is significantly influenced by image appearance variations. d Biological processes cause tissue structure differences, while sample preparation and scanning introduce notable appearance variations. Preserving pathological tissue structural differences is crucial for capturing biological features in AI learning for the purpose of diagnosis, but it’s important to minimize appearance variations that don’t relate to biological traits. Image appearance migration addresses this by migrating images from different scanners into a standard space. e Initially, mean RGB value distributions from various scanners may cluster distinctly due to appearance differences. f After applying image appearance migration, the data points converge into a compact, unified appearance, demonstrating the method’s effectiveness in standardizing digital pathology images for AI analysis.