It’s a great pleasure to welcome Udaranga Wickramasinghe from EPFL, Lausanne, Switzerland at our lab for an invited talk!
Abstract: CNN-based volumetric methods that label individual voxels dominate the field of biomedical image segmentation. However, 3D surface representations of the segmented structures are often required for tasks like shape analysis. They can be obtained by post-processing the labeled volumes which typically introduces artifacts and prevents end-to-end training. In this talk, I introduce Voxel2Mesh, a novel architecture that goes from 3D image volumes to 3D surfaces directly without any post-processing and with better accuracy than current methods when using smaller training datasets. I will discuss in detail about the motivation, design choices, strengths and limitations of the architecture. I will also discuss how this can help to accelerate the adoption of deep learning techniques for shape analysis in medical imaging.
Short Bio: Udaranga Wickramasinghe is a PhD student at CVLAB – EPFL advised by Prof. Pascal Fua. His research focuses on 3D surface extraction from volumetric images and ways to introduce prior knowledge into deep neural networks. Prior to joining CVLAB, he completed his master’s degree in Computer Science at EPFL, Switzerland in 2017 and his bachelor’s degree in Electronics and Telecommunication Engineering at University of Moratuwa, Sri Lanka in 2014.
Links & References
Voxel2Mesh on arxiv: https://arxiv.org/abs/1912.03681
Voxel2Mesh code: https://github.com/cvlab-epfl/voxel2mesh
Heart segmentation: https://arxiv.org/abs/2102.07899
Deep Active Surface models: https://arxiv.org/abs/2011.08826
This video is released under CC BY 4.0. Please feel free to share and reuse.