Fig. 1From: Automatically detecting Crohn’s disease and Ulcerative Colitis from endoscopic imagingExperimental workflow. After an upstream processing of the input images (blocks “Data splitting”, “Class balancing”, “Transform”), for each classification task three pre-trained ResNet variants of increasing complexity (“Model i”, \(i = 1,2,3\)) are used as weak learners and fine-tuned in 5-fold cross-validation on the input data (see Methods). A meta-model is then built by stacking ensemble of the three weak learners, evaluating the performance on the external test set in terms of different classification metrics (see Methods). N–P, negative versus positive; UC–CD, Ulcerative Colitis versus Crohn’s Disease; UC–N, Ulcerative Colitis versus negative; MCC, Matthews Correlation Coefficient (MCC); TNR, true negative rate; TPR, true positive rate; NPV, negative predictive value; PPV, positive predictive valueBack to article page