CLU-CNNs: Object detection for medical images

Publisher:
Elsevier
Publication Type:
Journal Article
Citation:
Neurocomputing, 2019, 350, pp. 53-59
Issue Date:
2019-07-20
Filename Description Size
1-s2.0-S0925231219305521-main.pdfPublished version1.38 MB
Adobe PDF
Full metadata record
Medical images have different characteristics from normal images. As an important feature, there usually exists data distribution difference between source domain and target domain for data scarcity and privacy. In this paper, a domain adaptation framework called CLU-CNNs is proposed, which is designed for medical images. CLU-CNNs uses ANCF and BN-IN Net to improve domain adaptation capability without specific domain adaptation training. Based on probability distribution assumptions of networks’ output, ANCF is a new path for domain adaptation. And BN-IN Net is embedded in fully convolutional networks to improve stability. This work has three key contributions: (1) A new object detection domain adaptation method is proposed in this paper without specific domain adaptation training. (2) Designed for medical images, CLU-CNNs performs well on small dataset, and is easy to be expanded. (3) CLU-CNNs obtains high positioning accuracy and fast speed when there is data distribution difference between source domain and target domain. Test on REFUGE CHALLENGE 2018, our way achieves state of the art performance.
Please use this identifier to cite or link to this item: