Impact of Skin Lesion Segmentation Labeling Errors on Convolutional Neural Networks

Accurate segmentation of skin lesions is critical for dermatological diagnosis and treatment. This study investigates the impact of label errors on the performance of convolutional neural networks in skin lesion segmentation. We conducted experiments using ground truth and simulated erroneous labels to train models and analyzed the effects of the errors using regression analysis. To simulate errors, we assume that errors are more likely to occur in more complex border regions. In addition, we add parameters to bias the errors towards either false positives or false negative regions. The overall effect of the false positive and false negative errors on the Dice coefficient is deceptively low. However, when we observe the false negative or false positive rate of the predictions, we see that they are significantly influenced by the bias of the label errors. In addition, label false positives have a more severe impact on prediction results than label false negatives. We recommend focusing on minimizing label false positives and suggest moving beyond mean Dice Score Coefficients for model evaluation.<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">1</sup><sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">1</sup>All of the training and analysis code is available on github.com/marinbenc/label_error_analysis.