In short, Stefan Klein's research entails the following:
Automatic grading and phenotyping of soft-tissue tumors through machine learning to guide personalized cancer treatment
'Soft tissue tumors (STT) are a rare and complex group of lesions with a broad range of differentiation. All STT subtypes greatly differ in their clinical behavior, aggressiveness, molecular background, and preferred treatments given. Diagnosis of the correct phenotype, the grade of aggressiveness, and molecular make-up is therefore of utmost importance. Diagnosis of STT is generally supported by imaging, such as computed tomography (CT) and magnetic resonance imaging (MRI). However, visual assessment by a radiologist tends to be subjective and not precise. Quantitative, computational (“radiomics”) imaging features and state-of-the-art Artificial Intelligence (AI) techniques based on machine learning could enable more objective and precise STT diagnosis. With the support of the Hanarth Foundation, we aim to develop a comprehensive STT diagnostic model, both for phenotyping and grading. This model will be trained and validated in a large, multi-center cohort, and evaluated in a clinical setting. The model will be based on quantitative image analysis by radiomics and deep learning. We hypothesize that, by considering multiple STT phenotypes at once instead of training a specialized model for each subtype, breakthroughs will be achieved with regard to the diagnostic performance of the AI model and its generalizability. Our AI model will guide diagnosis and treatment decisions, thereby facilitating personalized medicine.'
Brief summary of progress / results 2022
Within this project we aim to develop an artificial intelligence (AI) model supporting soft-tissue tumor diagnosis, by predicting tumor phenotype and grade based on MRI and/or CT imaging. In the first year of this project, we have focused on:
- development of a roadmap with clinically relevant diagnostic questions that might be tackled with AI;
- data collection;
- development of automated methods for tumor segmentation.
Regarding the last point, segmentation is prerequisite of many follow-up analyses by artificial intelligence methods. However, manual segmentation is a time-consuming process. We have therefore investigated the use of minimally interactive segmentation methods. One in which the clinician only is required to annotate a few points within the tumor, as well as a self-supervised learning approach in which the training of the segmentation model is initialized by a model trained to solve an auxiliary task for which more data is available.