Using AI to improve the positioning of 3D landmarks: correcting the biases of traditional registration methods

In this study, a 2D deep learning (DL) architecture was developed to automate the placement of 10 selected landmarks on 463 mouse skulls. These landmarks were chosen for the diversity of their locations and definitions. For each landmark, the predicted positions were compared with those obtained using registration approaches such as ALPACA. Euclidean distances relative to the manual ground truth, as well as the preservation of individual shape, were compared between the two approaches. The .tif images and associated manual landmarks are provided in the data attached to this metadata record. Furthermore, the weights of the AI models and the R codes used to perform the statistical analyses in the associated publication are also provided and described.