Analysis of Trained Convolutional Neural Network using Generative Adversarial Network
Analysis of Trained Convolutional Neural Network using Generative Adversarial Network
复制标题
DOI：
--
复制
发表时间：
2021
期刊：
影响因子：
--
通讯作者：
Yasuyuki Tsutsui;Yuki Shinomiya;Shinichi Yoshida
中科院分区：
文献类型：
--
作者：
Yasuyuki Tsutsui;Yuki Shinomiya;Shinichi Yoshida
近年来，由于医生短缺而增加的医生负担已经成为医疗场所中的问题，并且可以通过图像识别实现高度准确的诊断的基于CNN的方法作为自动化诊断以减轻医生负担的方法正在引起关注。但是，在医疗现场，需要向患者提供解释，基于CNN的方法尚未引入现场，因为很难解释结果。因此，我们提出了一种使用注意力引导的CycleGAN作为分析CNN识别的区域和模式的差异的方法，用于解释使用CNN的诊断，并从转换结果，焦点区域的相似性和使用CNN的诊断转换的有效性来评估分析的有效性。结果表明，Attention-Guided CycleGAN可以有效地用于CNN的分析，因为转换结果符合心脏肥大和无症状患者之间的症状，并且对于CNN的诊断也是有效的。
In recent years, the increasing burden on doctors due to the shortage of doctors has become a problem in the medical site, and CNN-based methods that can realize highly accurate diagnosis by image recognition are attracting attention as a method to automate diagnosis to ease the burden on doctors. However, in the medical site, it is required to provide explanations to patients, and CNN-based methods have not been introduced to the site because it is difficult to explain the results. Therefore, we proposed a method using Attention-Guided CycleGAN as a method to analyze the differences in regions and patterns recognized by CNN for the explanation of diagnosis using CNN, and evaluated the validity of analysis from the results of the transformation, the similarity of regions of focus, and the effectiveness of transformation for diagnosis using CNN. The results show that Attention-Guided CycleGAN may be effective for the analysis of CNN, because the transformation results are in line with symptoms between cardiomegaly and asymptomatic patients, and be also effective for diagnosis by CNN.