The clinical aim is to improve treatment response prediction of patients suffering from DLBCL based on information provided by FDG-PET in combination with clinical data. As 40% of the patients do not respond to standard chemotherapy, there is an urgent need to better predict treatment outcome to avoid futile treatments and severe side effects without clinical benefit.
Investigating different machine learning models and apply them to predict treatment response based on radiomics data obtained from FDG-PET scans.
Designing and training convolutional neural networks to automatically segment lesions from the scans and exploring the impact of using different loss functions
Combining the previously designed CNN segmentation model with the machine learning predictive model to generate a fully automate treatment prediction pipeline.
Developing a novel CNN which willl extract the predictive information directly from the scans. The robustness of this AI model will be tested using a test-retest study and different reconstruction settings to assess its clinical applicability.