Physics- and biology-informed deep-learning super-resolution imaging of cellular networks in tooth dentin

This project aims to improve current imaging capacity of microscopic cellular porosity in tooth dentin, a biomedical challenge with strong implications for fundamental biology in dentistry. This porosity has very specific geometry and topology, but shares some common features with the cellular porosity in bone or vascular systems: it is tubular in shape, organized on at least 2 distinct length scales and forms a network.

Deep-learning super-resolution is an active branch of image restauration with very a broad range of applications, including biomedical ones. Many super-resolution models have been proposed, with increasingly complex and larger architectures in a quest for precision and generalizability. However, it is very unclear precisely what drives performance of those generic models.

Our goal is therefore to use our extensive knowledge of the physics and biology of cellular porosity in dentin as leverage to design more efficient and explainable models for this specific type of applications. This mostly involves:

1. selecting and optimizing appropriate model architectures and loss-functions
2. identify the correct lengths scales for data sampling to train the models
3. designing adapted image quality metrics for model assessment. 

This project is strongly centered on AI for Computer Vision, but is very interdisciplinary by nature, with extensions in Physics of Complex Systems, Optical Microscopy and Imaging and Biophysics.