Research project Neural networks used for delineating tumors and risk organs in radiation therapy are not limited to, as is done today, only report which tissue is in each pixel. With the help of statistical methods, it is possible to obtain information about how certain the network is of its conclusions. This uncertainty information can be used to provide feedback on the quality of the delineations, which can form the basis for further training of the network, and be used to create more robust treatment plans.
The purpose of this project is to investigate how AI-generated information about uncertainties can be used to improve radiation therapy. The focus is on three areas: (1) How well can we quantify uncertainties in delineations and how can this information be used to inform when manual adjustments are needed? (2) Can the data generated during manual adjustments be used to continuously improve the networks that create the delineations in an effective way? (3) How can the uncertainties from the networks be integrated into the development of robust radiation treatment plans?
Background
Radiation therapy is an important part of the fight against cancer and is used in about half of all cancer cases. Its aim is to direct radiation at cancer cells while minimizing damage to other organs in the body as much as possible. An important step in planning radiation therapy is marking tumors and sensitive organs, which has traditionally been very resource intensive. With the advent of artificial intelligence (AI) in healthcare, this is changing. Research in semantic segmentation has enabled the automatic delineation of tumors and risk organs, and software for this purpose is now available for use in healthcare. Accurate delineation of tumors and risk organs in medical images image is difficult. This means that both human and AI-based methods are always associated with uncertainties. If the uncertainties are large, there is a risk of missing the target with the radiation, which can lead to treatment failure or increased side effects. The AI-based methods open a new opportunity for individualized uncertainties through several methods that have been developed recently for uncertainty estimation in deep neural networks. How well these methods work for quantifying uncertainties in AI segmentations within radiation therapy is currently unknown, nor do we have a good understanding of how much the dose plans can be improved if reliable uncertainty information were available.
Goals
The goal of the project is to investigate how reliable and useful AI-based uncertainty estimates are for radiation therapy with focuses on three areas:
1. Develop and evaluate methods for AI-based uncertainty estimates, concentrating on how well they can quantify actual segmentation uncertainties.
2. Investigate how uncertainties can be used to automatically identify delineations that may need manual corrections and explore how these corrections can be effectively used to improve the networks.
3. Investigate how the calculated uncertainties can be used to generate more robust dose plans.
Significance
Correctly quantified uncertainties provide insights into how reliable automatic segmentations are, which can lead to better decisions and more reliable dose plans. Automatic identification of uncertain segmentations combined with manual corrections used to improve the automatic segmentations leads to the automatic systems improving over time. This is of great benefit to both patients and the clinic, where resources can be saved as the methods become better and better with minimal manual work. Integrating the uncertainties into the dose calculations enables more individualized dose plans to be used, which is expected to result in more effective treatments with fewer side effects.
