Segmentation of lung nodule in CT images based on mask R-CNN

Menglu Liu, Junyu Dong, Xinghui Dong, Hui Yu, Lin Qi

Research output: Chapter in Book/Report/Conference proceedingConference contribution

507 Downloads (Pure)

Abstract

Due to the low-quality of CT images, the lack of annotated data, and the complex shapes of lung nodules, existing methods for lung nodules detection only predict the center of the nodule, whereas the nodule size is a very important diagnostic criteria but is neglected. In this paper, we employed the powerful object detection neural network “Mask R-CNN” for lung nodule segmentation, which provides contour information. Because of the imbalance between positive and negative samples, we trained classification networks based on block. We selected the classification network with the hightest accuracy. The selected classification network was used as the backbone of the image segmentation network—Mask R-CNN, which performs excellently on natural images. Lastly, Mask R-CNN model trained on the COCO data set was fine-tuned to segment pulmonary nodules. The model was tested on the LIDC-IDRI dataset.
Original languageEnglish
Title of host publication2018 9th International Conference on Awareness Science and Technology, iCAST 2018
PublisherIEEE
Pages95-100
Number of pages6
ISBN (Electronic)978-1-5386-5826-0
ISBN (Print)978-1-5386-5827-7
DOIs
Publication statusPublished - 1 Nov 2018
Event9th IEEE International Conference on Awareness Science and Technology - Fukuoka, Japan
Duration: 19 Sep 201821 Sep 2018
http://www.design.kyushu-u.ac.jp/~icast/

Publication series

NameIEEE iCAST Proceedings Series
PublisherIEEE
ISSN (Print)2325-5986
ISSN (Electronic)2325-5994

Conference

Conference9th IEEE International Conference on Awareness Science and Technology
Abbreviated titleiCAST 2018
CountryJapan
CityFukuoka
Period19/09/1821/09/18
Internet address

Fingerprint

Dive into the research topics of 'Segmentation of lung nodule in CT images based on mask R-CNN'. Together they form a unique fingerprint.

Cite this