Author: Ke Lu, Chunhao Wang, Ruoxu Xia, Zhenyu Yang, Fang-Fang Yin, Chulong Zhang, Lei Zhang, Rihui Zhang, Jingtong Zhao, Haiming Zhu 👨🔬
Affiliation: Duke University, Duke Kunshan University, Medical Physics Graduate Program, Duke Kunshan University, The First People's Hospital of Kunshan 🌍
Purpose: The human brain’s spherical geometry offers unique opportunities for improving the segmentation of tiny and irregular anatomical structures. We hypothesize that representing the brain in spherical coordinates, using a shell-by-shell slicing approach, can better capture these complex structures. This study aims to develop a deep learning model with spherical convolutions to segment glioma enhancing tumor (ET) regions using multi-parametric MRI (MP-MRI).
Methods: A total of 369 glioma patients with a 4-modality MP-MRI protocol from the BraTS2020 dataset were included in this study. A spherical slicing technique was applied to the 3D brain volume as a set of 2D spherical slices (in azimuthal 0°–360° and polar 0°–180° angles) for each MR modality. Specifically, the 3D volumetric data were first encapsulated within a sphere, with the centroid of the brain volume serving as the sphere's center. Spherical surfaces were generated with radii ranging from 1 mm to 140 mm, incremented in 1 mm steps. A U-Net CNN defined on spherical coordinates was subsequently developed to segment ET regions from these spherical slices. The proposed model was implemented with an 8:2 training-to-test set assignment. Model performance was assessed in both the spherical and Euclidean domains including Dice, sensitivity, accuracy, and specificity.
Results: Preliminary results demonstrated the effectiveness of the proposed model. In the spherical domain, our model achieved an average Dice=0.691, sensitivity=0.778, accuracy=0.993, and specificity=0.999. In Euclidean space, our model showed an average Dice=0.734, sensitivity= 0.768, accuracy=0.996, and specificity=0.998.
Conclusion: The developed deep spherical convolutional model successfully segmented glioma ET regions using spherical-defined MP-MRI images. The proposed a method offers a novel and effective approach for modeling tiny and irregular anatomical structures, with promising applications in other medical imaging segmentation tasks.