Fig. 5
AI training with OpenSlide extracted original MRXS and DICOM converted images does not result in systematic performance improvements or reductions. (a) AI models (ResNet18, cTransPath, Virchow, and Virchow2) were trained on local SIP data with the original MRXS (OpenSlide extracted) or the DICOM converted images and tested on TCGA data for different resolutions and data set sizes. Exemplary results for the AUC are shown in (b–d). Results in (b) belong to the cTransPath model trained on ovarian data at a resolution of 0.25 mpp and set of 17 patients. Results in (c,d) belong to the Virchov model trained bladder data at a resolution of 0.25 and 1 mpp and a set of 17 patients.
