Artificial intelligence for medical image analysis in radiation oncology: tumor characterization, treatment planning, and individual risk assessment in soft-tissue sarcoma and prostate carcinoma patients

This thesis explores the application of artificial intelligence techniques, such as radiomics and deep learning, on medical imaging for cancer patients undergoing radiation therapy. The research demonstrated that, for patients with soft-tissue sarcoma, a radiomic model combining quantitative image analysis and machine learning can predict patient outcomes more accurately than qualitative assessments by radiologists. The Delta-Radiomics concept was tested analysing MRI studies before and after therapy for optimal prediction of treatment response. Furthermore, it was shown that deep learning can effectively differentiate between more aggressive and less aggressive tumors, a pathological property known as "tumor grading." Additionally, a specific type of neural network called U-Net was successfully trained to automatically contour soft-tissue sarcoma tumor volumes, facilitating AI analysis and potentially future clinical treatment planning. Lastly, the thesis successfully applied radiomics to enhance the computed tomography-based detection rate of lymph node metastases in prostate cancer patients.