Uncertainty calibration in deep ensembles for medical imaging

Study how well deep ensembles and MC dropout quantify uncertainty on out-of-distribution chest X-rays, and propose a recalibration method. This project sits within the deep learning and medical imaging activities of our group at ETH Zürich, and builds on several recent publications and ongoing collaborations with industry and other research groups.

• Review the state of the art in uncertainty and formulate a precise research question
• Design and implement the methodology described above
• Evaluate results rigorously, including ablations and comparison with baselines
• Document the work in a thesis and, where results allow, a conference paper

You will join weekly group meetings, work closely with a PhD student as day-to-day supervisor, and have access to our computing/lab infrastructure. We expect independent work, curiosity and good communication. The exact scope is adapted to a master/phd thesis of about 6 months.

• Supervision by Anna Müller and an experienced team
• Help with applying for Erasmus+ or other mobility grants
• An international environment and the possibility to continue with a PhD for strong candidates

Apply through ThesisAbroad with your CV, transcript and a short motivation explaining why this topic interests you. Shortlisted candidates are invited to a 30-minute video call.

• Enrolled in a master or PhD-track programme in Artificial Intelligence or a related field
• Solid background in deep learning
• English (B2 or higher)
• Programming or lab experience relevant to the topic
• Ability to spend the full duration on site