Elvis Chun-Sing Chui1, Lawrence Chun-Man Lau1, Xin Ye1, Lik-Hang Hung2, Ka-Bon Kwok3, Kwoon-Ho Chow1, Ronald Man-Yeung Wong1, Edmond Wing-Fung Yau4, Sheung-Wai Law1, Wing-Hoi Cheung1* and Patrick Shu-Hang Yung1*
1Department of Orthopedics and Traumatology, The Chinese University of Hong Kong, Hong Kong
2Department of Orthopedics and Traumatology, Prince of Wales Hospital, Australia
3Department of Orthopedics and Traumatology, Alice Ho Miu Ling Nethersole Hospital, Hong Kong 4Koln 3D Technology (Medical) Limited Company, Science Park, Hong Kong
Applications of deep learning models and Convolutional Neural Network (CNN) have been achieving good performance in 3D medical image analysis. In this study, we present a fast and efficient 3D femur segmentation method based on V-net with a gradient pre-processor. Instead of feeding CT data into a standard V-net model, our proposed model is fed with gradient data based on CT scans using a gradient pre-processor, which forces the network to learn from the gradient field. Adopting an objective function based on dice similarity coefficient, the imbalance between the numbers of femur voxels against that of background could be addressed. A dataset of lower limb CT data with 60 samples was trained and tested on a pure V-net model and the proposed V-net model separately. Experimental results show that our proposed method could achieve better segmentation results (1.4% improvement in dice similarity coefficient) and a higher robustness as compared with the pure V-net model, which allows faster training speed and higher segmentation accuracy.
Deep Learning; Convolutional Neural Network; Volumetric Segmentation; V-net; Gradient pre-processor; Femur
Sing CC, Xin Y, Hang HL, Bon KK, Ho CK, Yeung WM, et al. Improving the Performance of V-Net Architecture for Volumetric Medical Image Segmentation by Implementing a Gradient Pre-Processor. Clin Surg. 2021; 6: 3383..