Automated diagnosis of melanoma

High technology solutions to the difficult task of selecting and monitoring moles (pigmented skin naevi) may be useful to keep accurate records of people’s skin. Adopting military surveillance and warfare technology,1 there are computer algorithms that search for changes in moles’ appearance over time. Deep convolutional neural networks analysis can group them into benign or malignant lesions with high accuracy.2 In a study by Esteva and colleagues,2 the convolutional neural networks algorithm differentiated between benign, malignant or non-neoplastic lesions with about 72% accuracy compared with about 66% accuracy by two dermatologists; for melanocytic lesions, the algorithm had a better sensitivity and specificity performance compared with the average of 21 dermatologists, although these findings still need to be replicated in independent datasets. Despite recent advances, there are still questions about how Australians can benefit from this technology and how it is best integrated into clinical practice.

Cancer agencies worldwide do not recommend screening for melanoma, but instead ask people to make skin self-examinations a habit and present to a doctor with moles of concern — although informal screening is widespread in Australia. Apps that provide easy access to personalised risk estimation may alert people to engage in such exams more frequently. Moreover, apps that guide people through the skin self-examination process may also be useful, as most people find this task complex.3 Once people notice a spot or mole, they may seek a clinical skin examination. Evidence that clinical skin exams are beneficial comes from the Queensland melanoma case control study4 and other similar studies that show that they lead to the detection of thinner melanomas. There are many apps that allow people to take and send photos of moles, but these are highly variable in sophistication and costs. Whether such technology is best placed in front of (for filtering out clearly benign lesions) or after a clinician’s diagnosis (for additional validation) is also matter of debate. Apps should not distract from the patient–doctor relationship, as the final decision about excision requires face-to-face consultations.

While technology solutions are promising, validation studies have mostly been small, have lacked a control group or have not been replicated in clinical practice. Independent big research initiatives, such as the International Skin Imaging Collaboration Challenge on Skin Lesion Analysis towards Melanoma Detection,5 are underway to take the momentum further. This healthy competition may be just what is needed to take the last steps to eradicate melanoma.