Automated Orientation Detection of 3D Head Reconstructions from sMRI Using Multiview Orthographic Projections: An Image Classification-Based Approach.

Recent studies in neuropsychiatry have highlighted the correlation between facial and brain dysmorphologies. One way of simultaneously analysing the brain and the face of a subject is by reconstructing a whole-head 3D model from structural magnetic resonance imaging (sMRI). However, the use of different reconstruction protocols generates undesired orthogonal rotations of the 3D models. This is a likely situation in multicentric studies that hampers the combination of data from different centers. Although the original sMRI files contain the subject orientation, it is not always possible to access this data. To solve this issue, in this work we propose a novel method to estimate the orientation of 3D heads with rotations of 90 $$^\circ $$ or multiples thereof around any of the three Cartesian axes as a required step for generating a normalised dataset in terms of orientation. Our proposal creates 2D images from orthogonal projections of the 3D object, transforming orientation estimation into an image classification problem. Experimental results show that our method, using three orthographic views of the 3D head to create the projection image and ResNet50 for classification, achieves an accuracy of 99.7%, which corresponds to 0.15 mean absolute error in rotation, outperforming state-of-the-art point cloud registration methods like DeepBBS and PRNet.