Fig. 2: Robustness to scanner variations.

a ROC curves of MSIntuit performance on MPATH-DP200 and MPATH-UFS cohorts. To compare performance on the exact same set of patients, we kept the subset of patients that passed QC on the two sets of slides (n = 536), and obtained an AUROC of 0.88 on both scanner, b Correlation of the predictions on the same slides on the UFS/DP200 scanners resulting in Pearson’s correlation of 0.98 (two-sided t test p < 0.001), c Prediction distribution for 30 slides, where each slide was digitised 8 times with the UFS scanner. Fleiss’ Kappa of 0.82 was obtained, showing an almost perfect agreement of the tool between the different digitisation of the same slide. d Heatmaps showing MSI score for each 112 × 112 μm tile for one representative slide digitised with two scanners, e Correlation of tile MSI scores on DP200 and UFS scanner. MSIntuit outputs a score for each tile, hence we also analysed the concordance of tile scores for a subset of 20 slides digitised with the two scanners (n = 272,527 tiles). A Pearson’s correlation of 0.92 was obtained (two-sided t test p < 0.001). The colormap representing the spatial density of points indicates that most tile scores were close to the diagonal, showing that tile scores were highly concordant. Source data are provided as a Source Data file.