Shape Segmentation

SPHARM-Net is a spherical harmonics-based convolutional neural network for vertex-wise inference. A rotation-equivariant convolutional filter can avoid the need for rigid data augmentation, and rotational equivariance can be achieved in spectral convolution without requiring a specific neighborhood definition. However, the limited resources of modern machines enable only band-limited spectral components that may result in the loss of geometric details. SPHARM-Net seeks to use a constrained spherical convolutional filter that supports an infinite set of spectral components. The proposed filter encodes all spectral components without relying on the full harmonic expansion to capture geometric details. The use of rotational equivariance significantly reduces training time for vertex-wise inference. The proposed convolution is fully composed of matrix transformations, offering efficient and fast spectral processing. While initially tested on brain data, SPHARM-Net has the potential to be extended for processing any spherical data.

• Leveraging Input-Level Feature Deformation with Guided-Attention for Sulcal Labeling

  Seungeun Lee, Seunghwan Lee, Ethan Willbrand, Benjamin Parker, Silvia Bunge, Kevin Weiner, Ilwoo Lyu

  IEEE Transactions on Medical Imaging, 44(2), 915–926, 2025

• SPHARM-Net: Spherical Harmonics-based Convolution for Cortical Parcellation

  Seungbo Ha, Ilwoo Lyu

  IEEE Transactions on Medical Imaging, 41(10), 2739–2751, 2022

We adapted spherical convolutional neural networks designed for generic semantic segmentation tasks to our sulcal labeling task. However, directly using the generic networks for our task is challenging due to two main reasons: limited training samples that capture anatomical variability and a lack of hierarchical neuroanatomical association. To adapt the generic networks to our specific problem, we propose spherical data augmentation and context-aware training. These techniques allow us to efficiently utilize existing training samples and accept a wide range of individual variability by considering training data synthesis and contextual information.

• Labeling Lateral Prefrontal Sulci using Spherical Data Augmentation and Context-aware Training

  Ilwoo Lyu, Shunxing Bao, Lingyan Hao, Jewelia Yao, Jacob Miller, Willa Voorhies, Warren Taylor, Silvia Bunge, Kevin Weiner, Bennett Landman

  NeuroImage, 229, 117758, 2021

• Automatic Labeling of Cortical Sulci using Convolutional Neural Networks in a Developmental Cohort

  Lingyan Hao, Shunxing Bao, Yucheng Tang, Riqiang Gao, Prasanna Parvathaneni, Jacob Miller, Willa Voorhies, Jewelia Yao, Silvia Bunge, Kevin Weiner, Bennett Landman, Ilwoo Lyu

  IEEE International Symposium on Biomedical Imaging (ISBI) 2020, IEEE, 412–415, Iowa City, IA, 2020

• Cortical Surface Parcellation using Spherical Convolutional Neural Networks

  Prasanna Parvathaneni, Shunxing Bao, Vishwesh Nath, Neil Woodward, Daniel Claassen, Carissa Cascio, David Zald, Yuankai Huo, Bennett Landman, Ilwoo Lyu

  Medical Image Computing and Computer Assisted Intervention (MICCAI) 2019, LNCS11766, 501–509, Shenzhen, China, 2019 · early accept, travel award