BA/MA: Shape Analysis on Meshes or Point Clouds from Medical Data

Advisor: Prof. Bjoern Menze
Supervision by: Anjany Sekuboyina
Fernando Navarro

Abstract

In this project we explore the feasibility of employing geometric deep learning or point cloud-based deep learning for extracting useful features from shapes in medical images.

Figure 1: Shape Analysis in different medical datasets [5]

In this project, we explore the feasibility of employing geometric deep learning or point cloud-based deep learning for extracting useful features from shapes in medical images.

Background and Motivation

Anatomical shape or morphology is an important pointer in population studies. „How does this sample differ from the rest?“. On the other hand, shape is not apparent, especially in 3D medical imaging. Moreover, typical analysis algorithms are image-centered, with no explicit processing of shape information. Therefore, in this project, we side-step image analysis, and move to alternate representations of 3D objects in the form of meshes and point clouds. We check the feasibility of learning useful information from these representations, which eventually can be used for augmenting the learning from images.

Tasks

Two directions of exploration according to student interest and skills, both with the option of the data type (meshes or point clouds):

• Fracture detection in vertebrae
• Brain tumor grading.

Prerequisites

The student should have prior knowledge in :

• Image processing
• Experience with deep learning
• Good programming skills (python, git)
• Experience with Pytorch
• Basics in Geometric DL (desired)
• Generative modeling
• Docker

Contact

Anjany Sekuboyina Fernando Navarro Bjoern Menze

References

[1] SplineCNN^?: Fast Geometric Deep Learning with Continuous B-Spline Kernels. Matthias Fey, Jan Eric Lenssen, Frank Weichert, Heinrich Müller

[2] Probabilistic Point Cloud Reconstructions for Vertebral Shape Analysis. Anjany Sekuboyina et al.

[3] Discriminative and Generative Models for Anatomical Shape Analysis on Point Clouds with Deep Neural Networks. Benjamın Gutierrez-Becker, Ignacio Sarasua, Christian Wachinger.