Fig. 2
Overview of the three consecutive modules for facial recognition. The first module (unsupervised morphological reconstruction) reconstructs a 2D face from a PET volume, building on an approach previously developed to reconstruct a 3D mesh. For this purpose, all the PET data are binarized, nonspatially connected regions are removed. For each patient, a facial isosurface is reconstructed to deduce a 3D mesh, which is finally converted to a 2D representation of the patient’s face by using raycasting methods. The second module (either with standard or deep learning-based approaches) denoises the 2D representation of the patients’ face. The third module performs face recognition thanks to landmark placement. To ensure the robustness of our analysis, we performed 5-fold cross-validation for the denoising and landmark detection analysis.
