Building representations of different brain areas through hierarchical point cloud networksDownload PDF

22 Apr 2022, 15:47 (edited 04 Jun 2022)MIDL 2022 Short PapersReaders: Everyone
  • Keywords: Neuroanatomy, Point Cloud, Deep Learning, PointNet, Brain Parcellation
  • TL;DR: We propose a point cloud-based deep learning method that processes the microstructures inside the images for brain mapping.
  • Abstract: Understanding how the microstructure varies across different brain regions is critical for disease modeling and brain registration. However, current deep learning approaches that work on image data directly may unintentionally focus on textures or other sources of noise in the data and fail to capture meaningful information about the underlying microstructures of interest. In this work, we propose a deep learning method that aims to build salient representations of microstructures inside neuroimage data by working on point cloud representations. We developed a hierarchical PointNet to process extracted 3D point clouds of brain anatomy to solve a brain region classification task. We validate our method on a micron-scale neuroimaging dataset, where we generated point clouds from both pixel-level segmentations and simple edge detection methods. In both cases, we show that point cloud-based models achieve better stability and performance when compared to 3D convolutional networks trained on the same brain region classification task. Our results in using “noisier” data from simple filtering operations provides initial evidence that point cloud representations could be a lightweight and data-efficient approach for brain parcellation. Keywords: Neuroanatomy, Point Cloud, Deep Learning, PointNet, Brain Parcellation
  • Registration: I acknowledge that acceptance of this work at MIDL requires at least one of the authors to register and present the work during the conference.
  • Authorship: I confirm that I am the author of this work and that it has not been submitted to another publication before.
  • Paper Type: novel methodological ideas without extensive validation
  • Primary Subject Area: Unsupervised Learning and Representation Learning
  • Secondary Subject Area: Learning with Noisy Labels and Limited Data
  • Confidentiality And Author Instructions: I read the call for papers and author instructions. I acknowledge that exceeding the page limit and/or altering the latex template can result in desk rejection.
1 Reply

Loading