by Manuela Braun

Scientists Julia Fligge-Niebling and Sireesha Chamarthi from the German Aerospace Center (Deutsches Zentrum für Luft- und Raumfahrt; DLR) are not physicians. They will never meet a patient in a doctor's office or hospital and have to decide whether a mole could be indicative of skin cancer. And yet, they will help to ensure that images of suspicious skin lesions can be recognised more reliably in the future, helping doctors during screening. The two researchers at the DLR Institute of Data Science are working with the German Cancer Research Center (Deutsches Krebsforschungszentrum; DKFZ) to train artificial intelligence (AI) systems so that they are able to evaluate images and assess the likelihood of cancer, adapted to the specific conditions in medical practices and to individual patients. The project, which is funded by the Helmholtz Association of German Research Centres, has the somewhat unwieldy name of 'Patient-specific diagnostic AI-systems via one-shot domain adaptation' (PSDAI). Now, the search has begun for a method that enables a trained AI to be adapted using very small data sets.

More than 230,000 people are diagnosed with skin cancer every year in Germany, and early detection offers the best chance of a cure. ABCDE is the formula that doctors use to first assess whether moles and birthmarks could be cancerous. A for asymmetrical, B for boundary, which in cancer can be uneven or blurred, and C for colour, which is concerning if the colour of a mole is unevenly distributed. D stands for diameter, which should not exceed five millimetres, and E stands for evolution, when a skin change suddenly becomes more pronounced. "The detection of skin cancer using artificial intelligence methods has been the subject of research for a long time," says mathematician Julia Fligge-Niebling, who heads the Machine Learning Group at the DLR Institute of Data Science.

Scientists Sireesha Chamarthi and Julia Fligge-Niebling are adapting an AI system to individual conditions in skin cancer screening. (Credit: DLR)

From the research lab to the hospital

Studies have already shown that artificial intelligence can be a useful aid in diagnosis. Systems with AI are certainly capable of detecting and classifying abnormalities. The challenge here: "Every hospital and every practice has different lighting conditions and cameras when recording suspicious skin areas. And every patient's skin colour and hairiness are unique," explains data scientist Sireesha Chamarthi. Even if the AI system has been trained to classify with tens of thousands or hundreds of thousands of data items, in individual cases it could be prone to errors because it is not precisely adjusted to match the relevant conditions. If neural networks from research labs are to be used in practices and hospitals in the future, they will have to become even smarter – or rather, they will have to be equipped with an algorithm that they can adapt in the shortest possible time and with only one or very few images.

Training for adaptation

DLR is concentrating on two methodologies – 'domain adaptation', adapting a trained AI model to datasets that are different from but related to a target dataset, and 'few-shot learning', which is used to train an AI model using a limited dataset. The DKFZ provides the training images for this and will also test the created algorithm with dermoscopic images and cell sections of skin lesions. "Training an AI-based diagnostic system is a lot more complex than we first thought," says Roman Maron, a computer and data scientist at the DKFZ. And yet studies have shown that AI already is on a par with the best doctors when it comes to skin cancer detection under laboratory conditions.

Moles instead of astronomical data

The project, which started in January 2022, is still in its infancy. In order to research and apply the 'few-shot learning' method, a neural network must first be trained with tens of thousands of images. Then the real research work begins; the method must be developed so that the AI can adapt to individual cases, such as doctors who photograph skin lesions with brighter light or different types of cameras, patients with many or very few dark hairs on their skin, and patients with very pale skin. 

Sireesha Chamarthi worked with astronomical data at the Indian Institute of Astrophysics in Bengaluru before joining DLR in 2020. The various classifications for skin cancer detection are new territory for her. "But that makes no difference to my work. You do need to understand the features with each new task – but it ultimately comes down to the right method."