An Automated CNN-based 3D Anatomical Landmark Detection Method to Facilitate Surface-Based 3D Facial Shape Analysis. R Huang, M Suttie, JA Noble.

Date: October 2019.
Source: Greenspan H. et al. (eds) Uncertainty for Safe Utilization of Machine Learning in Medical Imaging and Clinical Image-Based Procedures. CLIP 2019, UNSURE 2019. Lecture Notes in Computer Science, vol 11840. Springer, Cham.
Abstract: Maternal alcohol consumption during pregnancy can lead to a wide range of physical and neurodevelopmental problems, collectively known as fetal alcohol spectrum disorders (FASD). In many cases, diagnosis is heavily reliant on the recognition of a set of characteristic facial features, which can be subtle and difficult to objectively identify. To provide an automated and objective way to quantify these features, this paper proposes to take advantage of high-resolution 3dMD facial scans collected from a high-risk population. We present a method to automatically localize anatomical landmarks on each face, and align them to a standard space. Subsequent surface-based morphology analysis or anatomical measurements demands that such a method is both accurate and robust. The CNN-based model uses a novel differentiable spatial to numerical transform (DSNT) layer that could transform spatial activation to numerical values directly, which enables end-to-end training. Experiments reveal that the inserted layer helps to boost the performance and achieves sub-pixel level accuracy.
Article: An Automated CNN-based 3D Anatomical Landmark Detection Method to Facilitate Surface-Based 3D Facial Shape Analysis.
Authors: Ruobing Huang, Michael Suttie, J Alison Noble. Institute of Biomedical Engineering, University of Oxford, Oxford, UK | Nuffield Department of Women’s and Reproductive Health University of Oxford, Oxford, UK.