TY - GEN
T1 - Modified Hybrid Task Cascade for Lung Nodules Segmentation in CT Images with Guided Anchoring
AU - Xu, Shihuai
AU - Lu, Huijuan
AU - Ye, Minchao
AU - Yan, Ke
AU - Zhu, Wenjie
AU - Jin, Qun
N1 - Funding Information:
This work is supported by the National Natural Science Foundation of China under Grants No.61272315, No.61701468, and the Natural Science Foundation of Zhejiang Province under Grants No.LQ20F030015 and the outstanding student achievement cultivation program of China Jiliang University (2019YW16).
Publisher Copyright:
© 2020 ACM.
PY - 2020/2/15
Y1 - 2020/2/15
N2 - As lung cancer continues to threaten human health, Computer-Aided Diagnostic (CAD) plays an increasingly significant role in lung cancer diagnosis, and convolutional neural networks (CNNs) have shown the outstanding performance in image segmentation. In this work, Hybrid Task Cascade (HTC) is used to segment lung nodules that are difficult to find in CT images. Considering that lung nodules are usually quite small, this study integrates Feature Pyramid Network (FPN) into ResNet-50 to make full use of multi-scale feature and improve the segmentation accuracy of small target nodules. In addition, given that existing defects in Region Proposal Network (RPN), which refers to most of generated anchors are irrelevant to target objects, and the conventional method are unaware of the shapes of target objects, this work proposes to use Guided Anchoring to replace RPN in HTC and generate anchors more effectively. Experimental results on the LIDC-IDRI dataset demonstrate that the modified HTC improves the segmentation accuracy of lung nodules.
AB - As lung cancer continues to threaten human health, Computer-Aided Diagnostic (CAD) plays an increasingly significant role in lung cancer diagnosis, and convolutional neural networks (CNNs) have shown the outstanding performance in image segmentation. In this work, Hybrid Task Cascade (HTC) is used to segment lung nodules that are difficult to find in CT images. Considering that lung nodules are usually quite small, this study integrates Feature Pyramid Network (FPN) into ResNet-50 to make full use of multi-scale feature and improve the segmentation accuracy of small target nodules. In addition, given that existing defects in Region Proposal Network (RPN), which refers to most of generated anchors are irrelevant to target objects, and the conventional method are unaware of the shapes of target objects, this work proposes to use Guided Anchoring to replace RPN in HTC and generate anchors more effectively. Experimental results on the LIDC-IDRI dataset demonstrate that the modified HTC improves the segmentation accuracy of lung nodules.
KW - CT images
KW - Guided Anchoring
KW - Hybrid Task Cascade
KW - lung nodules segmentation
UR - http://www.scopus.com/inward/record.url?scp=85085929812&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=85085929812&partnerID=8YFLogxK
U2 - 10.1145/3383972.3384073
DO - 10.1145/3383972.3384073
M3 - Conference contribution
AN - SCOPUS:85085929812
T3 - ACM International Conference Proceeding Series
SP - 433
EP - 438
BT - Proceedings of the 2020 12th International Conference on Machine Learning and Computing, ICMLC 2020
PB - Association for Computing Machinery
T2 - 12th International Conference on Machine Learning and Computing, ICMLC 2020
Y2 - 15 February 2020 through 17 February 2020
ER -
