Detecting skin cancer with a smartphone app

Researchers have developed an app that photographs suspicious skin changes and sends them to hospital-based dermatologists for image analysis

Can artificial intelligence help dermatologists assess melanoma risk?

Translate from the original German language version

According to the German Cancer Society, more than 200,000 people in Germany develop skin cancer every year, and 3,764 people died from it in 2017 alone. The most dangerous form of skin cancer is the malignant melanoma. However, melanoma can also be cured very well if detected in time. The survival rate after five years is more than 95%. This makes early diagnosis and prevention all the more important.

The Fraunhofer Center for Assistive Information and Communication Solutions AICOS in Porto and Lisbon has now developed a solution that significantly accelerates the path to diagnosis. The Derm.AI solution combines smartphone photos of the skin spot with image analysis software and artificial intelligence. It provides a quick initial assessment of potentially dangerous skin changes. Dermatologists use this platform as a decision-making aid and then first analyse those cases with an increased risk of skin cancer.

Standardised photos are analysed by skin cancer specialists

The solution also aims to support the process of tele-dermatology in Portugal's healthcare system. "The problem of detecting skin cancer early has been a recurring theme among GPs in recent years. People who notice dark patches or other noticeable changes on their skin need a diagnosis quickly. But in regions where there are only a few specialists, it often takes longer to get an appointment for an initial examination. In addition, patients often have to travel long distances. This is where our Derm.AI solution comes in," explains Maria Vasconcelos, head researcher at Fraunhofer AICOS.

In the first step, the general practitioner takes a photo of the suspicious spot on the skin with the smartphone. The Fraunhofer team has developed an app for precisely this purpose. It runs on Apple's iPhone as well as on Android smartphones. The app helps to align the smartphone camera correctly and ensures that the photos are taken in the right resolution, with the right image detail and at the right distance. Two photos are taken, one as a close-up of the suspicious spot and one from a greater distance in order to also get the surroundings in the picture.

In this way, standardised images are created with the same settings in resolution, image detail, brightness and contrast. "The specialists can compare these images well with others and analyse them reliably," says Vasconcelos.

Image analysis with AI for prioritisation

The images created in the GP practice are then sent via the internet to the dermatology ward of a hospital. Now, software equipped with artificial intelligence enters the scene. It analyses the photos of the suspicious spot, compares them with reference data and the data of other patients and then estimates the risk of skin cancer: The spot in question is then marked as "harmless", "risky" or "dangerous". This is not yet a diagnosis, but merely an initial assessment. This serves to prioritise the order in which the cases are viewed.

The physicians then first take those cases for which the AI software indicates a higher probability of malignant skin cancer and which therefore need to be diagnosed quickly. "The software does not make its own decision, but only a preselection based on probabilities. The actual examination and diagnosis is always in the hands of the dermatologist in charge," Vasconcelos explains.

After analysing images and patient data such as age, gender or previous illnesses, the dermatologist in the hospital can start a consultation with the responsible general practitioner via telephone or video conference or schedule an appointment to examine the patient directly.

Examining suspected skin cancer cases on site

In cases where the dermatology specialists are not sure, a quick appointment is made for a personal on-site examination. There, the affected skin area is examined under a reflected light microscope, for example, or tissue samples are analysed in a biopsy to gain certainty.

About 80% of the cases in which patients present themselves at the general practitioner's office with suspicious skin changes turn out to be harmless birthmarks or moles after image analysis and consultation between the general practitioner and the dermatologist. The GP can then quickly give the all-clear. And the patients save themselves long waiting times and a trip to the hospital for an appointment.

In the case of patients whose skin changes are not clearly harmless or point to the less dangerous light skin cancer, the physician asks them to come back in three months, for example, to have a new photo taken of the suspicious spot.

Deep-learning software with GP expertise

The AICOS researcher and her team developed the algorithm for image analysis in the Derm.AI project. The deep-learning software was fed with image data and information from 4,000 patients. In the subsequent prioritisation of cases by the algorithm, the medical expertise of dermatologists was also incorporated. "We had many discussions with GPs and dermatologists to understand what they really need. The feedback from the physicians on Derm.AI is very good," Vasconcelos is pleased to say.

Partners in the project are Portugal's Ministry of Health, the Oncology Hospital of Coimbra, the University Hospital of Porto and the Health Centre of the city of Guarda. Currently, the AICOS researchers are analysing the results from the practical use of the smartphone app. They are also refining and optimising the deep learning model of the AI software.