Medical image segmentation through learning on unstructured 3D data

Abstract

Medical image segmentation is an essential task for diagnosis and treatment monitoring. However, deep learning methods used to perform this task generally have a relatively high computation cost, particularly when the processed images are large volumes containing millions of voxels. This PhD thesis therefore proposes to replace the original voxel representation by a lighter point cloud representation in order to develop more efficient segmentation algorithms. In the first part of this thesis, we propose a workflow that extracts a contour point cloud from the original image and feeds it to SCONet, our implicit network adapted from convolutional occupancy networks for the task of multi-organ segmentation. We evaluate our method on two abdominal CT datasets. In the second part, we provide an in-depth study of the influence of the point cloud quality on SCONet’s performance and identify which points are more useful for segmentation. In the third part, we evaluate the robustness of our method to domain shifts and find that SCONet performs well when confronted to dataset shifts and to changes in imaging modality.

Jury

HEINRICH Mattias, Professor, University of Lübeck, Rapporteur 

NAEGEL Benoît, PU, Université de Strasbourg, Rapporteur 

DESVIGNES Michel, PE, Université Grenoble Alpes, Examinateur 

DIGNE Julie, DR, Université Claude Bernard Lyon 1, Examinatrice 

MATEUS Diana, PU, Centrale Nantes, Examinatrice 

VALETTE Sébastien, CR, INSA Lyon, Directeur de thèse 

KÉCHICHIAN Razmig, MCF, INSA Lyon, Co-directeur de thèse