Computer-aided polyp detection system powered by deep neural networks has achieved high performance but also suffers from data insufficiency. To address this problem, recent researches focus on synthesizing new colonoscopic images by Generative Adversarial Network (GAN). However, the synthesized images follow the same distribution as that of the training dataset, which limits the performance of the detectors re-trained on it. Recent studies show that adversarial examples can expand the data distribution and thus adversarial training can effectively improve the robustness of deep neural networks. Inspired by these two factors, this paper proposes a data augmentation framework to directly produce false negative colonoscopic images via GAN and the adversarial attack. The synthesized polyps are natural and experiments on three popular detectors show that compared with using GAN alone, producing false negative images by the adversarial attack can further improve the performance of the re-trained detectors.
基金:
National Natural Science Foundation of China [61977046]
语种:
外文
被引次数:
WOS:
第一作者:
第一作者机构:[1]Shanghai Jiao Tong Univ, Inst Image Proc & Pattern Recognit, Shanghai, Peoples R China
通讯作者:
通讯机构:[1]Shanghai Jiao Tong Univ, Inst Image Proc & Pattern Recognit, Shanghai, Peoples R China[2]Shanghai Jiao Tong Univ, Inst Med Robot, Shanghai, Peoples R China
推荐引用方式(GB/T 7714):
He Fan,Chen Sizhe,Li Shuaiyi,et al.COLONOSCOPIC IMAGE SYNTHESIS FOR POLYP DETECTOR ENHANCEMENT VIA GAN AND ADVERSARIAL TRAINING[J].2021 IEEE 18TH INTERNATIONAL SYMPOSIUM ON BIOMEDICAL IMAGING (ISBI).2021,1887-1891.doi:10.1109/ISBI48211.2021.9434050.
APA:
He, Fan,Chen, Sizhe,Li, Shuaiyi,Zhou, Lu,Zhang, Haiqin...&Huang, Xiaolin.(2021).COLONOSCOPIC IMAGE SYNTHESIS FOR POLYP DETECTOR ENHANCEMENT VIA GAN AND ADVERSARIAL TRAINING.2021 IEEE 18TH INTERNATIONAL SYMPOSIUM ON BIOMEDICAL IMAGING (ISBI),,
MLA:
He, Fan,et al."COLONOSCOPIC IMAGE SYNTHESIS FOR POLYP DETECTOR ENHANCEMENT VIA GAN AND ADVERSARIAL TRAINING".2021 IEEE 18TH INTERNATIONAL SYMPOSIUM ON BIOMEDICAL IMAGING (ISBI) .(2021):1887-1891