Lung Cancer Classification Using Masked Autoencoder Pretrained on J-MID Database

Ren Tasai, Guang Li, Ren Togo, Minghui Tang, Takaaki Yoshimura, Hiroyuki Sugimori, Kenji Hirata, Takahiro Ogawa, Kohsuke Kudo, Miki Haseyama

Abstract: Annotation for lung cancer classification using chest CT images typically requires specialized knowledge and is time-consuming and costly. Additionally, it has been reported that in self-supervised learning, accuracy improves when the datasets used for pretraining and downstream tasks belong to the same domain. In this paper, we propose a self-supervised learning method based on a Masked Autoencoder pretrained on the high-quality J-MID database, which is in the same domain as the downstream task. Experimental results show that our approach outperformed previous methods pretrained on ImageNet, which consists of a large number of natural images.