MATTO-GBM Multimodality Artificial intelligence open-source

Our research aims to improve the diagnosis and treatment of a type of brain cancer called glioblastoma (GBM). Due to a higher rate of local recurrence (LR), these tumours are challenging to treat and, therefore, have poor chances of survival. One of the main reasons for high LR is that the treatment procedures for GBM have remained almost unchanged for decades, especially in radiotherapy (RT) treatment. During RT planning, difficulty in identifying the properties and the precise location of the tumour (segmentation) can contribute to treatment failure. Currently, doctors use magnetic resonance imaging (MR) to visualize the tumour borders. However, the scans obtained from an MR machine cannot differentiate between tissue changes caused by RT (pseudoprogression) and modifications due to tumour growth. Trying to remedy the pseudoprogression with additional radiation treatment could lead to the death of healthy body tissues. On the other hand, positron emission tomography (PET) imaging is good at distinguishing between pseudoprogression and actual tumour growth. Therefore, imaging using complementary PET/MR systems offers an all-in-one solution with superior diagnostic power while minimizing time consumption, and patient discomfort. Although the impact of PET/MR imaging in RT planning has been substantial, its implementation in the clinical workflow is still challenging. The first challenge is that, in daily RT practice, the radiation oncology physicians manually perform tumour segmentation, which is a time-consuming process, and sometimes doctors may have differing views (bias). The second challenge in daily RT practice is that the delivered radiation dosage is calculated based on the tissue densities measured by computed tomography (CT). Therefore, physicians combine and register (aligning multiple medical images to a standard system) the segmentations of organs at risk and tumours (obtained from PET/MR images) and tissue densities (obtained from CT). The registration of images is crucial to determine the dosage locations and limits. But, registering images captured separately using CT and PET/MR systems could lead to errors. Hence, CT synthetization from MR would be beneficial to reduce registration errors. We will train supervised AI algorithms to segment GBM from images (which helps reduce bias and time consumption). Furthermore, generative AI algorithms will be developed by us to synthesize CT from MR (thus reducing registration errors). These could benefit the patient by helping to deliver the appropriate amount of radiation dose during RT. In addition, we will develop models that could predict the time and location of the recurrence of GBM. All the resultant software will be combined in an open-access tool, allowing the integration of our results by different health institutions worldwide to adapt GBM treatment based on the individual risk pattern. In conclusion, our work proposes an essential step in personalized medicine options for GBM.