Single-Valued Risk Estimation for Segmentation Including Data Uncertainty

In recent decades, the field of machine learning, especially image segmentation, has attracted immense interest from the general public and in research. However, the safety analysis, necessary for use in safety-critical systems, introduces new challenges as the model behaviour is often not explicitly designated. Consequently, existing white-box testing methods are not feasible, leaving black-box testing as a viable alternative. This, however, leads to non-deterministic bounds, which only hold at a certain confidence level, making them unable to be used in classical risk assessment techniques like fault trees, Markov models, or reliability block diagrams. To overcome this problem, we present a new method to obtain single-valued upper bound estimation for the risk in segmentation. The latter can be applied with a wide range of loss functions. The bound can be obtained with a pre-computable additive constant and the measured test performance. Hence, it is straightforward for any practitioner to use and allows deployment on existing segmentation models. Furthermore, it can be seen and used as an automated selection of the confidence level. Consequently, existing risk assessment techniques can be applied again to estimate risk in the early stages of system design.