- Granted applications
- 2022 Call
- 2021 Call
- 2020 Call
- 2019 Call
- Before 2019
- Application 2023
- Contact us
Artificial Intelligence in Oncology - Supporting scientific research
MAASTRO Clinic, Maastricht University Medical Centre+
Recent radiology studies on head/neck cancers suggest that a sizeable proportion of patients receive invasive neck node dissection surgery after chemo-radiotherapy, but actually derive no survival benefit from the extra treatment. Making this judgement is presently very difficult for doctors, and likely depends strongly on the radiologist’s skill and the biological patterns of each person’s cancer. But in patients at high risk of the cancer returning, the additional surgery is known to be beneficial. This dilemma suggests a real need for a clinical decision support tool to help doctor make better decisions. Artificial intelligence (AI) is especially good for detecting hidden patterns in radiology images, so this project will use AI to exploit the wealth of imaging data collected in top cancer hospitals worldwide. This will allow us to identify subtly different patterns in tumours and simulate the results of different treatments, so that doctors in future can recommend an individually personalized treatment and reduce the burden of unnecessary neck surgery.
This project uses large amounts of patient data and clinical images of head and neck cancer, distributed around the world, to try to reduce the number of invasive surgeries after chemo-radiotherapy treatment. Artificial intelligence (AI) helps us here, because we can “teach” an AI to search for subtle signs that are hard for human eyes to see, and also the AI can examine the images for us instead of exchanging patient data between researchers. This is known as Privacy-preserving Federated Deep Learning. To date, we have re-designed tools for making private data Findable-Accessible-Interoperable-Reusable to be more easily deployed, remotely-supported and scaled up to a large number of clinics; this is an extremely important step due to the effect of Covid-19 on travelling and working face-to-face.
Next, we consulted with clinicians and legal experts from over 20 clinics in 11 countries to make an easily re-usable ethics/legality framework that is compliant with patient privacy laws all over the world. We have given this to the Dutch Health-RI Personal Health Train initiative, and our framework is already being re-used in two new projects (with many more to come). Lastly, we have made good progress towards correcting errors in state of the art AI tools and making it privacy-secure for federated deep learning. We are right now robustly testing this AI within a large multi-institutional global consortium, and will begin collection of new patient data internationally during 2022.
To date, this Hanarth-funded project has demonstrated open source and disease agnostic federated deep learning architecture at global scale, similar to fully centralized learning. The benefit of federated learning is that patients' data does not need to be transferred between hospitals or across borders, thus increasing
privacy safeguards. For data to be used in this privacy-sensitive way, it has to be made FindableAccessible-Interoperable-Reusable (FAIR), thus we have prepared open source tools that helps doctors share their data in a safe method. The focus is now to increase the sophistication of "deep radiomics" models that search medical images of head and neck cancer in greater detail, to look additional information that will help us to predict how likely it is that a cancer would return after treatment.