[4Yin2] Interactive session 2

[4Yin2-38] Building image classification models using a limited number of chest X-ray images

〇Isao Sano¹, Keisuke Tsukada¹, Masashi Nakatomi¹, Yasuo Sugitani¹ (1.Chugai Pharmaceutical Co., Ltd.)

Keywords:ChestX-ray8, lightweight-GAN, Medical data

The development of accurate diagnostic imaging technology is one of the major issues in medicine. Traditionally, in the development of medical imaging technology utilizing machine learning, it has been difficult to construct a high-precision model because of the limited data that can be used for the learning. However, the recent development of various generation models suggests the possibility of constructing high-precision models based on a limited number of image data. This study used a generative adversarial network to generate data with disease characteristics from limited public data of chest X-ray images (NIH ChestX-ray8), constructed image classification models using these datasets, and examined the effect of this method on model accuracy. Common thoracic lesions such as atelectasis and cardiomegaly were included in this study. The results suggested that accurate image classification models can be constructed with limited data.