Fig. 5: Cross-center performance enhancement of metastasis cancerous tissue detection with data privacy-preserving.

A WSI-level prediction probability maps of single-center pathology models for detecting metastatic cancer tissue across centers. Top row: The first two columns are the original WSI from Center 2 in CAMELYON17 and the corresponding expert-annotated mask. The third column shows the detection result of the raw patches classifier trained in Center 1 of CAMELYON16 using a CNN. The fourth column shows the detection probability map of the patches classifier with HSV augmentation when training. The fifth column shows the result of using the StainLUT trained in Center 1 of CAMELYON16 for real-time preprocessing and then detecting with the patches classifier. The last column shows the detection result with HSV augmentation added when training the patches classifier and StainLUT model stain normalization added when testing. Bottom row: Corresponding WSI of the red-boxed region at the top row. Where the probability maps of model detection and the original WSI image are fused and visualized. B Quantitative results of F1 Score and IoU for different schemes predicting cancerous tissue probability values greater than 0.5. C Quantitative AUC evaluation of different training schemes for binary classification of extracted image patches from WSI. Left: Model training in Center 2 of CAMELYON16 and same-center testing in Center 2 of CAMELYON17. Right: Model training in Center 1 of CAMELYON16 and cross-center testing in Center 2 of CAMELYON17. Experiments are repeated five times independently.