Algorithm for image recognition for early detection of Malignant Melanoma

Important results from the project

The project aim was, together with Karolinska Hospital and KTH, to develop and implement an AI decision support tool for diagnosis of melanoma. The goal was to develop an algorithm based on dermatoscopic images, and then test the algoritm in a clinical set-up. A model of an image recognition algorithm was developed and integrated. An important part of the project was to sort and structure data and images to develop the algorithms. Also, to plan for next stage of development.

Expected long term effects

An algorithm with 90% accuracy was developed and implementation of this in Dermicus platform was carried out. The clinical flow of developed algorithms was tested through the platform and verified that the technology works. But it was also found that more flexible presentation of the results is required for it to be fully usable. E.g. support the skin consultants by providing diagnostic suggestions with probability of diagnosis. The project has led to a better understanding and what is needed in future in algorithm development and data quality.

Approach and implementation

WP1 was a preparation phase where the requirements for the algorithms were specified and the data structured. Under WP2 conducted development, sorting of data, implementation and testing. WP3 was the analysis and summary phase. An important aspect of getting AI algorithms with high accuracy is having a structured data and diagnostic verified images. Structuring of data has therefore been ongoing during a larger part of the project and it provided a better basis for further development of the algorithms. Results has been presented at, eg. Vitalis 2019.