Anatomical consistency and confidence estimation of cardiac segmentation of ultrasound images using variational auto-encoders
Artificial intelligence shows great promise in the field of medicine. Indeed, neural networks can learn to perform many tasks that would otherwise take hours for physicians to accomplish. For example, neural networks can learn to classify each pixel in cardiac echography images with respect to the anatomical region (cardiac echography segmentation). This allows for faster diagnostic of various pathologies. However, neural networks are not infallible, and it is not trivial to identify these failure cases without expert intervention. The goal of our project is to propose a new confidence estimation technique to validate predictions made by neural networks for segmentation. This validation will be tailored to the task of cardiac echography as particular characteristics of this type of imaging, such as time dependency between predictions, will be considered.