TY - GEN
T1 - A Coarse to Fine Framework for Multi-organ Segmentation in Head and Neck Images
AU - Pu, Yan
AU - Kamata, Sei Ichiro
AU - Wang, Youjie
N1 - Publisher Copyright:
© 2020 IEEE.
PY - 2020/8/26
Y1 - 2020/8/26
N2 - Radiotherapy is widely used in the treatment of head and neck cancer. Due to the harmfulness of radiation, it is necessary to protect our healthy organs during the radiotherapy. Therefore, the accurate delineation of diseased region and surrounding healthy organs is the precondition for doctors to make the radiation plan. In real life, the delineation work is usually done manually. It is time-consuming and requires high professional skill. A fast and accurate organ segmentation method can greatly improve the efficiency of treatment. Most CT image datasets are 3D volumes and each volume can be divided into a series of 2D slice images. For multi-organ segmentation task, how to generate the stable organ features from CT images is still the plagued problem. For 2D framework, which processes the images slice by slice, the network cannot learn the correlation between continuous slices. It will lead to the loss of spatial information. For 3D framework, which processes the images volume by volume, the patch training is commonly used to against the massive increase of network parameters. The 3D patch will limit the maximum reception field of the network. For the organ, which is larger than the patch size, it is easy to lose global information. To solve these incompatible problems, we proposed a coarse to fine framework to take advantage of both 2D framework and 3D framework. The multi-view coarse network is designed to generate the organ probability maps and the coarse segmentation mask in 2D case. The organ volumes are extracted with the probability maps. These organ volumes are sent to the organ-based fine network to refine the mask of each organ in 3D case. Our proposed method is tested on the Head and Neck Automatic Segmentation Challenge datasets in 2015 and predict for 9 different organs. The result show that our framework performs the lowest error range for most organs and three of them achieve the top evaluation results in comparison with existing methods. Contribution-The main contribution of this paper is to propose a novel two-stage, Coarse to Fine, framework for multi-organ segmentation and verify its effectiveness in head and neck CT images.
AB - Radiotherapy is widely used in the treatment of head and neck cancer. Due to the harmfulness of radiation, it is necessary to protect our healthy organs during the radiotherapy. Therefore, the accurate delineation of diseased region and surrounding healthy organs is the precondition for doctors to make the radiation plan. In real life, the delineation work is usually done manually. It is time-consuming and requires high professional skill. A fast and accurate organ segmentation method can greatly improve the efficiency of treatment. Most CT image datasets are 3D volumes and each volume can be divided into a series of 2D slice images. For multi-organ segmentation task, how to generate the stable organ features from CT images is still the plagued problem. For 2D framework, which processes the images slice by slice, the network cannot learn the correlation between continuous slices. It will lead to the loss of spatial information. For 3D framework, which processes the images volume by volume, the patch training is commonly used to against the massive increase of network parameters. The 3D patch will limit the maximum reception field of the network. For the organ, which is larger than the patch size, it is easy to lose global information. To solve these incompatible problems, we proposed a coarse to fine framework to take advantage of both 2D framework and 3D framework. The multi-view coarse network is designed to generate the organ probability maps and the coarse segmentation mask in 2D case. The organ volumes are extracted with the probability maps. These organ volumes are sent to the organ-based fine network to refine the mask of each organ in 3D case. Our proposed method is tested on the Head and Neck Automatic Segmentation Challenge datasets in 2015 and predict for 9 different organs. The result show that our framework performs the lowest error range for most organs and three of them achieve the top evaluation results in comparison with existing methods. Contribution-The main contribution of this paper is to propose a novel two-stage, Coarse to Fine, framework for multi-organ segmentation and verify its effectiveness in head and neck CT images.
KW - CT images
KW - Head and neck
KW - Organs segmentation
UR - http://www.scopus.com/inward/record.url?scp=85099879467&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=85099879467&partnerID=8YFLogxK
U2 - 10.1109/ICIEVicIVPR48672.2020.9306647
DO - 10.1109/ICIEVicIVPR48672.2020.9306647
M3 - Conference contribution
AN - SCOPUS:85099879467
T3 - 2020 Joint 9th International Conference on Informatics, Electronics and Vision and 2020 4th International Conference on Imaging, Vision and Pattern Recognition, ICIEV and icIVPR 2020
BT - 2020 Joint 9th International Conference on Informatics, Electronics and Vision and 2020 4th International Conference on Imaging, Vision and Pattern Recognition, ICIEV and icIVPR 2020
PB - Institute of Electrical and Electronics Engineers Inc.
T2 - Joint 9th International Conference on Informatics, Electronics and Vision and 4th International Conference on Imaging, Vision and Pattern Recognition, ICIEV and icIVPR 2020
Y2 - 26 August 2020 through 29 August 2020
ER -
