Learn more about the use case:
FZI Research Center for Information Technology
AI-assisted diagnostics, Research project: Explainable AI
What was the mandate for the collaboration/development of an AI-based use case/solution/service?
The Intelligent Diagnostics project aims to improve the diagnosis of skin cancer. The diagnosis of skin cancer, by far the most common form of cancer, is currently not always reliable; 15% of tumors cannot be correctly diagnosed. Within the scope of the project, an AI-based diagnostic system for the early detection of skin cancer is to be developed with the aid of quantitative, multi- or hyperspectral imaging under structured irradiation and subsequent data processing with regard to 3D topology of the surface and other parameters relevant for diagnostics.
The interdisciplinary solution approach followed in the project covers several areas from the development of the optical measurement system, to the design and implementation of the AI infrastructure to collect and manage the data, to the design and training of the deep learning models.
What technologies and AI methods were used and why?
In the Intelligent Diagnostics project, neural networks are used to perform classification of skin lesions to diagnostic classes. For this purpose, the skin lesions are first imaged with a hyperspectral camera system, and the resulting hyperspectral images represent the inputs of the 3D faulting-based networks.
In addition, research is being conducted on Explainable AI methods to highlight regions in the images that are particularly important for the classification decision. This gives the physician the ability to understand which features in the hyperspectral images the AI system used to calculate the prognosis.
Another approach being considered as part of the project is federated learning. In future scenarios, this will make it possible to use measurement systems in several different skin clinics without having to send image data and thereby violate data protection guidelines, for example.
With which partners was the project implemented with which respective parts?
The Intelligent Diagnostics project is an innBW project involving the institutions FZI Research Center for Information Technology, ILM - Institute for Laser Technologies in Medicine and Measurement Ulm, Hahn-Schickard Villingen-Schwenningen, Hahn-Schickard Stuttgart, NMI Natural and Medical Sciences Institute at the University of Tübingen and, as a subcontractor, the University Hospital Tübingen.
How did you contribute to the transfer of knowledge and technology to the client?
Knowledge and technology transfer takes place via several channels. On one hand, press releases on the current state of development and research have been and are being published. On the other hand, the Intelligent Diagnostics project is and has been presented at various events. These include workshops such as "DIGInostik" or a dermatologists' workshop in Tübingen, but also the presentation of the project in conference papers and through posters. In addition, a project-supporting committee is informed about achieved results in events within the framework of the project.
A project website will provide a further opportunity to pass on acquired knowledge to interested companies and individuals.
In what timeframe did your company implement the AI solution?
The AI solutions will be implemented in the project from project month 4 to project month 17. This corresponds to a total of 14 months. In the initial phase, we will focus on the design and initial training of the neural network. The integration of Explainable AI approaches and the exploration of possible Federated Learning extensions will be done subsequently, in parallel with the further development of the neural network.
A conclusion to your AI solution:
The Intelligent Diagnostics project is currently in project month 13 and is not yet complete. The AI systems are currently trained on synthetically generated data. Among other things, these were created on the basis of simple RGB images. Results that are currently available may therefore differ from the final results. The current test accuracy (on the synthetic data) is 94%, which is already a partial success compared to other systems (accuracy of 85%).
The AI solution in the Intelligent Diagnostics project can access approximately 1 GB of data per hyperspectral image of a skin lesion. This amount of useful information per prediction gives hope that good results will also be achieved for predictions related to the natural images.
One final piece of advice/tip to other entrepreneurs looking to apply AI:
It is well known that AI systems need a sufficiently large amount of data to achieve good results. In the Intelligent Diagnostics project, the amount of data is severely limited, but good results are achieved with relatively little data. This is certainly due to the use of transfer learning. The neural network in the project is pre-trained with simple RGB images and synthetic hyperspectral images and only adapted to the natural data afterwards. An analogous approach is possible in many application fields and can lead to better results.
More info at: