The AI Challenge

Melanoma detection is hard, and spotting it early can save lives.

How can one spot the difference?

Even an expert eye can underestimate the difference in appearance that skin lesions present across age groups and ethnicities - What could look suspicious on an East-Asian 35-years-old could look normal on a 50-year-old Southern European.

The advantage AI can bring is that of condensing the knowledge of dermatologists across the world into the software.

This knowledge is known as training data, and is composed of images plus annotations that represent a doctor’s opinion on the mole. This can be in the form of a shape encompassing the outer shape of the mole (often ambiguous) to train computer vision algorithms, a reference to the histopathology scan, and tags representing its appearance or level of risk.

AI learns to compress this combined knowledge into neural networks, which can evoke memories of the training data when presented with a case that exhibits similar features to an example seen in the past.

The importance of Accurate AI Training Data

Curating the training data for any medical application is no small feat. Present too many benign lesions, and the AI will become too biased towards them.

Tag your demographics improperly, and the AI may ignore information that it cannot otherwise see or investigate, such as the age or ethnicity of the patient. V7’s technology is being used to apply tags on every patient and lesion to accurately represent both demographics and 16 lesion types.

Populations that go for mole scans are very imbalanced - with people from sunny parts of the world visiting dermatologists more frequently. It’s therefore crucial to ensure that the technology is trained on representative samples of these people.

In order to avoid other forms of bias, V7’s AI will semi-automatically spot every mole across the 3,000 images captured by iToBoS per patient, generate a cropped version, and human labelers will ensure that the segmentation around each mole is perfect and doesn’t include “distractors” such as items of clothing, tattoos, or jewellery that could bias the AI towards the wrong result (in machine learning this is known as an adversarial example).

As AI models within iToBoS are trained by data scientists within the consortium, V7’s software will use a workflow system to process every detected mole across multiple neural networks to combine their opinions, as well as dermatologists in case any of the moles are flagged as suspicious.

These dermatologists will also have access to previous images of the mole if available, to see if there have been any changes.

Skin lesions being semi-automatically segmented by human expert labelers leveraging AI models

In order to apply supervised learning to the problem, we need to see enough balanced data in order to catch the typical features of a melanoma across populations. Having annotated datasets is crucial and the deep learning revolution taught us that having computing power is not enough. We need curated datasets to start learning in an effective way.