Brain tumors represent only about 3,5% of all new cancers in adults but reduce life expectancy by more than 20 years on average. Brain cancer is the deadliest type of cancer in people under the age of 40 and one of the most common types of cancer in children. The tumours are classified into more than 150 different subtypes, from slow-growing, well-defined, benign tumours to diffuse malignant tumours with fast growth and very short life expectancy. While other cancer types give more stereotypical symptoms, brain tumors can give a wide range of symptoms depending on which brain structures and thus brain functions that are affected. The tumors are heterogeneous and there is great variability in aggressiveness, growth rate, and treatment responses.
Symptoms and prognosis vary widely and can be difficult to predict even for experienced doctors.Brain tumour patients may have a high symptom burden, both related to having incurable cancer and due to progressive neurological decline, seizures, and adverse treatment effects. Since clinical presentations, tumour locations, and growth patterns are all very diverse and often unpredictable, selecting optimal treatment course for any given patient is extremely difficult. Unfortunately, the recent progresses in both surgical and oncological multimodal treatment only benefit a smaller proportion of patients. Thus, personalized treatment is attractive for brain tumor patients as the natural course and treatment responses are highly variable ranging from patients that do not respond to therapy but only experience the risks to those who survive long enough to experience severe delayed adverse effects of treatment. Any potential therapeutic benefit on overall survival needs therefore to be judiciously balanced against potential treatment-induced side effects which may limit patients’ estimated quality of life, which would pave the way for personalized decision-making.
In this project, we aim to use machine learning to build computer models to predict the clinical outcome of new brain tumor patients considering at the same time the characteristics of the individual patient and the large amount of knowledge contained in the data from thousands of previous patients. We will apply machine learning and more specifically deep learning methods to radiological and clinical patient data, with a special focus on three main aspects. First, tools for quantitative radiological assessment of brain tumors including: (i) automatic segmentation and classification of pre-operative, intraoperative, post-operative, and follow-up MR/CT/US images to quantify initial tumor localization and volume, (ii) characterization of tumor resection during surgery, (iii) quantifying localization and volume of relevant structures postoperatively such as residual tumor, infarction, bleedings, resection cavity, or FLAIR hyperintensity, (iv) assessing tumor response to radiation/chemotherapy, and (v) quantifying tumor growth and regrowth. Second, the combined use of clinical and radiological data for personalized prediction of clinically important parameters such as survival, expected residual tumor and post-operative functional level such as Karnofsky score and quality of life. Finally, generative models to reconstruct clinically realistic data conditioned on patient-specific inputs and treatment scenarios, in order to address sparsity and irregularity of longitudinal data, and to generate future forecasting of tumor evolution or response.
