Fig. 1

Pipeline to obtain the best optimal after selecting gender, polyp size threshold, and feature transformation. The training data was used for the optimization process. First, the KDE bandwidth was optimized for each feature, then the feature that decreased the most the discrimination performance was eliminated, these steps were iterated until 3 features remained. The best 6 models from all the iterations were selected and evaluated on the test data to select the best of all. All the model selections were based on the MCC value