Home > Published Issues > 2019 > Volume 7, No. 3, September 2019 >

Effectiveness of Pseudo 3D Feature Learning for Spinal Segmentation by CNN with U-Net Architecture

Naofumi Shigeta, Mikoto Kamata, and Masayuki Kikuchi
School of Computer Science, Tokyo University of Technology, Tokyo, Japan

Abstract—In the medical field, automatic extraction of spinal region from CT images has been desired. Among various methods for image segmentation, one of the convolutional neural network models called U-Net [1] has been shown to attain good performance with small data set size. Previous study by Kamata et al. [2] applied U-Net for spine segmentation task and achieved 82.7% accuracy for unlearned CT images. However, the method had difficulty in the precision of the 3D shape. This study attempted extraction of spine region with higher precision by adopting pseudo 3D feature learning for U-Net.

Index Terms—medical image processing, spine segmentation, convolutional neural network, U-Net

Cite: Naofumi Shigeta, Mikoto Kamata, and Masayuki Kikuchi, "Effectiveness of Pseudo 3D Feature Learning for Spinal Segmentation by CNN with U-Net Architecture," Journal of Image and Graphics, Vol. 7, No. 3, pp. 107-111, September 2019. doi: 10.18178/joig.7.3.107-111