Comparison of convolutional neural network architectures for robustness against common artefacts in dermatoscopic images

Abstract

Introduction: Automated classification of dermatoscopic images via neural networks shows comparable performance to clinicians in experimental conditions, but can be affected by artefacts like skin markings or rulers. It is unknown whether specialized neural networks are equally affected, or more robust to artefacts. Objectives: Analyse robustness of three neural network architectures, namely ResNet34, Faster R-CNN and Mask R-CNN. Methods: We identified common artefacts in the public HAM10000, PH2 and the 7-point criteria evaluation datasets, and established a template-based method to superimpose artefacts on dermatoscopic images. The HAM10000-dataset with and without superimposed artefacts was used to train the networks, followed by analysing their robustness against artefacts in test images. Results: ResNet-34 and Faster R-CNN models trained on regular images perform worse than the Mask R-CNN models when tested on images with superimposed artefacts. Artefacts in all tested images led to a decrease in area under the precision-recall curve values of 0.030 for ResNet-34 and 0.045 for Faster R-CNN in comparison to only 0.011 for Mask R-CNN. However, changes in model’s performance only became significant with 40% or more of the images having superimposed artefacts. We could also show that loss in performance occurs when the training was biased by selectively superimposing artefacts on images belonging to a certain class. Conclusions: Instance segmentation architectures may be helpful to counter the effects of artefacts, and further research on related architectures of this family should be promoted. Our introduced template-based artefact insertion mechanism could be useful for future research.