Diagnostic accuracy of dermatoscopic image retrieval

Abstract

A neural network is a computer system modelled on the human brain. Neural networks can learn to detect visual elements such as colours, shapes and edges by themselves, and as such they have shown great improvements in automated classification and diagnosis of skin tumours. Unfortunately, in the most common format, they only provide final probability values as, for example “5% melanoma, 60% nevus, 25% seborrheic keratosis”. Although experimental results have shown networks to be fairly accurate, it is hard for a user to have insight into the prediction, and how to integrate it with their own decision process. One option to make a neural network “explainable” is to simply let it search for and present similar images on a database, much like looking up a similar case one has seen in a book – only faster and more comprehensive. This also has the upside that when images don't look anything close to the searched skin lesion, a user immediately knows not to trust the automated analysis in that case. The authors showed in this study, that if a network provides similar images, these images can convey the same accuracy as the conventional “5% melanoma, 60% nevus, 25% seborrheic keratosis” output. They further found that searching for up to only 16 similar images is enough to provide that information. This study suggests that simply searching for similar images can harness the prediction abilities of neural networks while being understandable for an end user.