바로가기  >

Abstract
The purpose of this study is implementation of an anthropomorphic model observer using a convolutional neural network (CNN) for signal-known-statistically (SKS) and background-known-statistically (BKS) detection tasks. We conduct SKS/BKS detection tasks on simulated cone beam computed tomography (CBCT) images with eight types of signal and randomly varied breast anatomical backgrounds. To predict human observer performance, we use conventional anthropomorphic model observers (i.e., the non-prewhitening observer with an eye-filter, the dense difference-of-Gaussian channelized Hotelling observer (CHO), and the Gabor CHO) and implement CNN-based model observer. We propose an effective data labeling strategy for CNN training reflecting the inefficiency of human observer decision-making on detection and investigate various CNN architectures (from single-layer to four-layer). We compare the abilities of CNN-based and conventional model observers to predict human observer performance for different background noise structures. The three-layer CNN trained with labeled data generated by our proposed labeling strategy predicts human observer performance better than conventional model observers for different noise structures in CBCT images. This network also shows good correlation with human observer performance for general tasks when training and testing images have different noise structures.

Affiliations

Minah Han  1 , Jongduk Baek  2
1 College of Engineering, Yonsei University, School of Integrated Technology, 162-1 Songdo-dong, Yeonsu-gu, Incheon, 406-840, Korea (the Republic of).
2 School of Integrated Technology College of Engineering, Yonsei University, 162-1 Songdo-dong, Yeonsu-gu, Incheon 406-840, KOREA, incheon, Korea (the Republic of).