Deep Learning in Medical Image Analysis

This seminar will discuss current research in the field of machine learning-based biomedical image processing. In contrast to general image analysis applications the medical domain provides special challenges that we want to focus on within the seminar:

• Data scarcity: It is rather common that research on complex medical applications faces the problem of only small amounts of available data. This is rarely due to intrinsic rareness of certain medical cases, but rather to difficulties related to the use of highly sensitive personal information, which is well-protected by law. Current research hence deals with approaches that get by with little or no annotated data at all.
• Robustness: Often decisive between life and death, algorithms in the medical domain necessarily need to ensure robustness as a criterion. Outliers have to be discovered automatically and treated separately during processing. In a more general sense, the processing systems should themselves be aware about the uncertainty in their computations and provide the user with related quantitative information.
• Generalizability: Medical applications are highly subject to variability. This includes, for example, different versions and settings of recording devices as well as different modes of handling by physicians. With the intention of broad applicability beyond a specific setting, solid generalizability of the method is required.

 A detailed list of topics will be released around the briefing date.