Nicolas - Roduit MS, PhD, Presenter: Nothing to Disclose

Rodolphe - Meyer MD, Abstract Co-Author: Nothing to Disclose

Philippe - Bijlenga MD,PhD, Abstract Co-Author: Nothing to Disclose

Antoine Geissbuhler, Abstract Co-Author: Nothing to Disclose

According to the face recognition literature, face features such as eyes, nose or mouth are particularly relevant for identifying a person. Thus, removing them is a mandatory task when exchanging radiological images within multi-site projects. In the European project @neuRIST (Integrated Biomedical Informatics for the Management of Cerebral Aneurysms) we have developed an algorithm making the patient’s face unrecognizable after 3D reconstruction from a set of images.

The very few existing depersonalization software tools are difficult to use, have high computational cost and may alter clinical information. On this base, we considered developing an algorithm fulfilling the following criteria: Preserving of the patient's identity integrity; Keeping the intra-cranial space completely unaltered to allow post-processing; Being robust against re-identification processes or reverse engineering attempts; Providing a fast and reliable automated process with low computational cost. Our multi-platform implementation is able to hide facial features in CT and MR images and to remove personal identifying text information found in DICOM headers, according to the DICOM standard recommendation. The 3D image depersonalization process consists in detecting automatically the sensitive face features and in hiding them by covering the face with a shapeless mask made of pseudorandom noise. This approach destructs nothing from the original images.  

Tests have shown that the 4 mentioned criteria were fulfilled. In particular, attempts to reverse the depersonalization process for retrieving a recognizable face have all failed. Globally, the quality of the results is excellent for CT images. Radiologists and clinicians were able to read, interpret and diagnose from all our images satisfactorily. Regarding few MR series some issues were noticed, mostly because MR images pixels are nosier than those of CT images and are not calibrated.

This permanent masking technique does not alter relevant medical information indispensable for diagnosis and research purposes.

This user-friendly application will facilitate data exchange between medical centers or for educational purposes by preserving the privacy of patients' data.

http://media.rsna.org/media/abstract/2009/8004936/8004936_7kcv.jpg

Roduit, N, Meyer, R, Bijlenga, P, Geissbuhler, A, Image Depersonalization: A New Efficient Approach.  Radiological Society of North America 2009 Scientific Assembly and Annual Meeting, November 29 - December 4, 2009 ,Chicago IL. http://archive.rsna.org/2009/8004936.html