Canonical orientation and segmentation of CT scans of skulls via 3D neural networks

Point-clouds have emerged as an essential tool, across various fields thanks to their flexibility. As a totally unconstrained data structure, they are used for many purposes, pooled together by the need of capturing, storing and processing 3D models. The use of CAD applications has become a mainstay in a variety of sectors, ranging from architectural design and construction industry to medical field, in particular in maxillofacial surgery, which is beginning to make increasing use of image-guided surgery, virtual surgical planning and the production of 3D printed anatomical models. The currently used approach, in surgical planning, is the direct manual handling of 3D models that requires lots of time and experts skilled in sculptor processing and in CAD modelling. The desired goal is to make this 3D modelling as automatic as possible. The objective of this thesis is about building a part of an overall pipeline, for a bigger research project, that is about the treatment of point-clouds, for a medical application, representing the outer shape of skulls. Specifically, the attention is focused on two steps of the overall 3D processing pipeline, to facilitate it, in the best possible way: the alignment of the skulls in a canonical pose and the separation of the points belonging to the cranium, from those ones of the mandible. The alignment step is dealt with a neural network that is fairly able to recognize the position of the skull in the space and move it, with the aim of placing it in the canonical position, with a rotation error expectation of less than 10◦ and few millimiters in translation. Regarding the segmentation phase, an ensemble of several models is employed, reaching the 99% of mean Intersection over Union. In the final step, high-resolution point clouds derived from segmentation are utilized to generate meshes for both the cranium and mandible. This pro- cess employs screened Poisson surface reconstruction, resulting in meticulously detailed meshes.