Reason for shared misclassifications by CNNs and dermatologists

Determine the reason underlying the shared misclassifications of certain dermatoscopic images that are consistently misclassified by multiple convolutional neural network models and also frequently misdiagnosed by expert dermatologists.

Background

The study identifies a subset of dermatoscopic images that are systematically misclassified by several CNN architectures and simultaneously difficult for expert dermatologists, with markedly reduced agreement metrics.

Although image quality partially explains these failures, the fundamental cause of these common misclassifications remains unresolved and is explicitly highlighted as an open question.