Medical Image Analysis
The Medical Image Analysis group conducts theoretical and applied research in image processing, computer vision, and artificial intelligence for the analysis of medical image datasets. The focus of our research relies on the paradigm of evidence-based image modeling and personalized medicine, aiming at effectively using medical image information, and image computing technologies, to leverage the understanding of diseases and medical conditions of the central nervous system, and to support and improve the healthcare system relying on the analysis of medical image information.
Accurate Quantification and Radiomics Analysis for Brain Lesions
Magnetic Resonance Imaging (MRI) and its variants are a powerful imaging modality that encompasses rich anatomical and physiological information at a high resolution. In neurosciences, these modalities have become a standard in clinical practice. However, the interpretation of the images requires the combined use of different modalities, which leads to the need of computer-assisted technologies. The group has developed several methodologies to analyze MRI images with focus on multimodal image segmentation for brain image lesion analysis studies. These developments are driven by clinical requirements such as computation speed, robustness, and use of standard clinical imaging protocols.
Accuracy is particularly paramount for an image-guided brain lesion quantification technology. Through a strong interdisciplinary collaboration with the department of neuroradiology, at the University Hospital, Bern, our interdisciplinary group has developed over the years accurate and clinically-relevant (i.e. in line with clinical requirements) solutions based on machine learning methodologies for automated brain tumor segmentation, stroke lesion segmentation, and multiple-sclerosis lesion segmentation, which have ranked among top-approaches at MICCAI (Medical Image Computing and Computer-Assisted Interventions) challenges, top-venue of the medical image computing field. Our seminal work on automated brain tumor volumetry was awarded the Young Scientist Publication Impact Award 2016, in recognition for being the most-impactful MICCAI work of the last five years, as well as the Ypsomed Innovation Award 2016.
Automated brain lesion quantification technologies are now used within the MANAGE project, for Multidimensional Response Assessment in Glioma Patients, which is an interdisciplinary effort aiming at developing longitudinal radiomics technologies and non-invasive biomarkers providing a better assessment of disease progression and patient response to therapy.
Uncertainty and Interpretability of Medical Image Segmentation Technologies using Deep Learning technologies
Next to accuracy, the robustness of computer-assisted technologies is fundamental for their effective deployment and integration in medicine. Particularly, it is crucial to develop technologies that can cope with computer errors stemming from the large heterogeneity of medical images, the complex pathophysiology of disease, among other factors. To this end, our group is developing algorithms that check the reliability of machine learning’s results by yielding uncertainty estimations of computer-generated results, which can be used to change the paradigm, so medical experts are no longer executioners of the task (e.g. brain tumor delineation), but use this information to monitor and correct them in a time-effective manner. In addition, as the amount of collected medical image information is rapidly growing, it is vital to develop Human-Machine Interfacing technologies (HMI) to ensure scalability of time-effective monitoring and correction technologies of computer-generated results.
Our group is researching methodologies to enhance the interpretability of machine learning models, so their decisions can be inspected (e.g. is the machine capturing the relevant relation in the data?), and interpreted by human (opening of the “black box”, e.g. If a system fails, why does it fail?). Enhancing interpretability of machine learning methods is essential in medicine, so to build trust with these systems, but it is also very important in line with discussions pointing to decision-making and a “right to explanation”.
Motivated by the current decoupling between the design of medical image sequences, and their exploitation through machine learning algorithms. In collaboration with MRI physicists from the academic and private sectors, our group is researching machine learning methodologies that are being applied at the image formation process stage, with the overarching goal of designing the best combination of MRI sequences and machine learning algorithms.
Towards Streamlined and High-throughput Data Curation Processes
Our group is establishing technologies for automatic quality assessment of curated data, as well as the reliability of the machine learning models produced with curated medical image information. On the one hand, automatic quality assessment of curated data is essential for high-throughput data curation of a highly heterogeneous and error-prone human interaction process of medical image information in the clinical routine. On the other hand, it is crucial to research and develop technologies that can inspect the reliability of machine learning models derived from this data.
During 2017, we initiated a Swiss-wide initiative to create infrastructure and technologies for a local and distributed radiomics platform, which features a data curation workflow occurring within the daily clinical routine. By leveraging the daily clinical workflow with human-machine intelligence technologies, we aim at creating a rich and sustainable symbiosis between their daily clinical needs, and the data curation process needed for biomedical research, therapy assessment (e.g. clinical trials), and in general for the improvement of data-driven biomedical engineering technologies