Machine Learning for Landmark Detection in Biomedical Applications

Machine Learning aims at developing models able to accurately predict an output variable given the value of some input variables by using a dataset of observed (input, output) pairs. In the recent years, the development of new Machine Learning algorithms as well as the increase of computing capabilities have made these methods very popular and successful to address various image processing related tasks.

One of these tasks is landmark detection, which consists in finding the coordinates of one or several interest points in images. Landmark detection finds many applications in computer vision. In this thesis, we focus on two of them, both related to bioimaging. The first is morphometrics, where landmark coordinates are used to measure the size and the shape of body parts. The second is image registration, where the coordinates of the landmarks are used to compute the deformation between two images.

During this thesis, we have developed an automated landmark detection algorithm combining tree-based machine learning models with multi-resolution pixel descriptors. Starting from an algorithm used for cephalometric landmark detection, we have progressively extended it in order to fit the needs of morphometric analyzes, where a wide variety of image datasets and body types are observed. We carefully analyzed the behavior of our algorithm in order to provide detailed insights about its performance on new image datasets. We then extended our landmark detection algorithm to 3D images and used it to perform CT-CBCT rigid registration. Finally, we studied the relevance of using post-processing steps based on the landmark shape structure given the specificities of biomedical applications.

Throughout this work, we evaluated our method on four different datasets: three datasets concerning 2D morphometrics, and one concerning 3D image registration. On these datasets, we showed that our algorithm could reach state of the art performance while providing additional genericity regarding its application on datasets containing different types of images.