講演抄録/キーワード |
講演名 |
2020-12-18 16:20
[ショートペーパー]Case Discrimination: Self-supervised Learning for classification of Medical Image ○Haohua Dong・Yutaro Iwamoto(Ritsumeikan Univ.)・Xianhua Han(Yamaguchi Univ.)・Lanfen Lin(Zhejiang Univ.)・Hongjie Hu・Xiujun Cai(Sir Run Run Shaw Hospital)・Yen-Wei Chen(Ritsumeikan Univ.) PRMU2020-64 |
抄録 |
(和) |
Deep Learning provides exciting solutions to problems in medical image analysis and is regarded as a key method for future applications. However, only a few annotated medical image datasets exist compared to numerous natural images. In this paper, we propose a model to investigate transfer learning by self-supervised learning using medical images. It is widely known that the results of Computerized Tomography (CT) scan are 3D volume images. There are lots of slices in CT or Magnetic Resonance Imaging scan images. So why not make these slices to a class? It is imperative to formulate this intuition as a self-supervised feature learning at the case-level. Our results of the experiment demonstrate that, under self-supervised feature learning settings, our method surpasses the transfer learning using ImageNet on classification. By experiment using unannotated dataset, our method is also remarkable for consistently improving test performance with a few annotated data. By fine-tuning the learned feature, we further obtain competitive results for self-supervised learning and classification tasks. |
(英) |
Deep Learning provides exciting solutions to problems in medical image analysis and is regarded as a key method for future applications. However, only a few annotated medical image datasets exist compared to numerous natural images. In this paper, we propose a model to investigate transfer learning by self-supervised learning using medical images. It is widely known that the results of Computerized Tomography (CT) scan are 3D volume images. There are lots of slices in CT or Magnetic Resonance Imaging scan images. So why not make these slices to a class? It is imperative to formulate this intuition as a self-supervised feature learning at the case-level. Our results of the experiment demonstrate that, under self-supervised feature learning settings, our method surpasses the transfer learning using ImageNet on classification. By experiment using unannotated dataset, our method is also remarkable for consistently improving test performance with a few annotated data. By fine-tuning the learned feature, we further obtain competitive results for self-supervised learning and classification tasks. |
キーワード |
(和) |
Self-supervised learning / 医用画像処理 / 画像分類 / / / / / |
(英) |
Self-supervised learning / Medical image processing / Image classification / / / / / |
文献情報 |
信学技報, vol. 120, no. 300, PRMU2020-64, pp. 151-155, 2020年12月. |
資料番号 |
PRMU2020-64 |
発行日 |
2020-12-10 (PRMU) |
ISSN |
Online edition: ISSN 2432-6380 |
著作権に ついて |
技術研究報告に掲載された論文の著作権は電子情報通信学会に帰属します.(許諾番号:10GA0019/12GB0052/13GB0056/17GB0034/18GB0034) |
PDFダウンロード |
PRMU2020-64 |