Fig. 2: Results of MSI and POLE prediction experiments with Vision Transformer based pipeline. | npj Precision Oncology

Fig. 2: Results of MSI and POLE prediction experiments with Vision Transformer based pipeline.

From: Deep learning for dual detection of microsatellite instability and POLE mutations in colorectal cancer histopathology

Fig. 2

Results are shown for external validation on TCGA (The Cancer Genome Atlas, International) and APHP Resection and Biopsy (Assistance Publique–Hôpitaux de Paris, France). Each patient is shown with 5 dots representing their prediction scores from 5-fold cross-validation. Highlighted dots correspond to the median AUROC model based on the TCGA cohort. The classification threshold is set to 0.5 for all cohorts. Mean and standard deviation of the predicted MSI score are shown, with values computed for driver mutations only in case of POLE/POLD1. Microsatellite status is indicated by color, wild type by WT, mutated by MUT, driver mutations by “d”, liver metastasis samples by green boxes and early-onset colorectal cancer patients (age at diagnosis <50 years)58 by a green triangle. Arrows point to the corresponding heatmaps of selected samples. a In the TCGA testing cohort, prediction scores were calculated for patients with MSI (excluding POLE/POLD1 mutations, top chart), MSS (excluding POLE/POLD1 mutations, middle chart) and POLE/POLD1 mutations (bottom chart). MSS group includes MSI-L and MSS. b, c In APHP resection (b) and biopsy (c) testing cohort, prediction scores were calculated for patients with MSI (excluding POLE/POLD1 mutations, top chart), MSS (excluding POLE mutations, middle chart) and POLE (bottom chart). d The mean and standard deviation for Predicted MSI Scores in panel a–c based split are calculated for the cohort and ground truth regarding microsatellite and POLE/POLD1 status. For POLE/POLD1 mutated cases, only driver mutations are considered. The icons indicating the origin of the cohorts are sourced from www.flaticon.com.

Back to article page