SINA: Sharp Implicit Neural Atlases by Joint Optimisation of Representation and Deformation

Medical image atlases simplify the analysis of large-scale datasets and are used for various applications, such as population analysis or probabilistic segmentation. They are often synthesized by minimising geometric distances. This formulation allows for fuzzy atlases, resulting in reduced interpretability and flexibility. Furthermore, most atlas generation methods are limited in terms of large anatomical variations and non-aligned data. We propose SINA (Sharp Implicit Neural Atlases), a novel framework for medical image atlas synthesis, leveraging the joint optimisation of data representation and registration. By iteratively refining sample-to-atlas registrations and modelling the atlas as a continuous function in an Implicit Neural Representation, we demonstrate the possibility of achieving atlas sharpness while maintaining data fidelity. Our approach is evaluated on 2D and 3D datasets of varying complexity against popular atlas synthesis methods. Our method ranks first regarding the synthesis of a complex chest X-ray atlas, second in the easier task of pre-aligned MRI images, and is capable of handling high-resolution abdominal CTs while composing representative and sharp atlases. Source code and trained models are published on GitHub .