Class Activation Maps for the disentanglement and occlusion of identity attributes in medical imagery

Abstract

Deriving patients' identity from medical imagery threatens privacy, as these data are acquired to support diagnosis but not to reveal identity-related features. Still, for many medical imaging modalities, such identity breaches have been reported. To cope with this, some de-identification methods based on the generation of synthetic data have been explored in the literature. However, in this paper, we try to perform, instead, an occlusion of the personal identifiers directly on the data by means of Class Activation Maps, in such a way that diagnosis related features do not get altered.