Comparison of thresholds for a convolutional neural network classifying medical images

Abstract

Our aim is to compare different thresholds for a convolutional neural network (CNN) designed for binary classification of medical images. We consider six different thresholds, including the default threshold of 0.5, Youden's threshold, the point on the ROC curve closest to the point (0,1), the threshold of equal sensitivity and specificity, and two sensitivity-weighted thresholds. We test these thresholds on the predictions of a CNN with InceptionV3 architecture computed from five datasets consisting of medical images of different modalities related to either cancer or lung infections. The classifications of each threshold are evaluated by considering their accuracy, sensitivity, specificity, F1 score, and net benefit. According to our results, the best thresholds are Youden's threshold, the point on the ROC curve closest to the point (0,1), and the threshold of equal sensitivity and specificity, all of which work significantly better than the default threshold in terms of accuracy and F1 score. If higher values of sensitivity are desired, one of the two sensitivity-weighted could be of interest.