このページは機械翻訳されています。他のページは英語で表示される場合があります。
View in English
オープンワールドの医療画像セグメンテーションのための汎用的な基盤モデルとデータベース
- Siqi Zhang 1, Qizhe Zhang 2, Shanghang Zhang 3, Xiaohong Liu 1,4, Jingkun Yue 1, Ming Lu 2, Huihuan Xu 1, Jiaxin Yao 1, Xiaobao Wei 2, Jiajun Cao 2, Xiang Zhang 5, Ming Gao 5, Jun Shen 5, Yichang Hao 6, Yinkui Wang 7, Xingcai Zhang 8,9, Song Wu 4, Ping Zhang 1, Shuguang Cui 10, Guangyu Wang 11
- Siqi Zhang 1, Qizhe Zhang 2, Shanghang Zhang 3
- 1State Key Laboratory of Networking and Switching Technology, Beijing University of Posts and Telecommunications, Beijing, China.
- 2State Key Laboratory of Multimedia Information Processing, School of Computer Science, Peking University, Beijing, China.
- 3State Key Laboratory of Multimedia Information Processing, School of Computer Science, Peking University, Beijing, China. shanghang@pku.edu.cn.
- 4South China Hospital, Medical School, Shenzhen University, Shenzhen, China.
- 5Department of Radiology, Sun Yat-Sen Memorial Hospital and Guangdong Provincial Key Laboratory of Malignant Tumor Epigenetics and Gene Regulation, Medical Research Center, Sun Yat-Sen Memorial Hospital, Sun Yat-Sen University, Guangzhou, China.
- 6Department of Urology, Peking University Third Hospital, Beijing, China.
- 7Key Laboratory of Carcinogenesis and Translational Research (Ministry of Education/Beijing), Gastrointestinal Cancer Center, Peking University Cancer Hospital and Institute, Beijing, China.
- 8Department of Materials Science and Engineering, Stanford University, Stanford, CA, USA.
- 9John A. Paulson School of Engineering and Applied Sciences, Harvard University, Cambridge, MA, USA.
- 10School of Science and Engineering (SSE), the Future Network of Intelligence Institute (FNii) and the Guangdong Provincial Key Laboratory of Future Networks of Intelligence, The Chinese University of Hong Kong, Shenzhen, China.
- 11State Key Laboratory of Networking and Switching Technology, Beijing University of Posts and Telecommunications, Beijing, China. guangyu.wang24@gmail.com.
- 0State Key Laboratory of Networking and Switching Technology, Beijing University of Posts and Telecommunications, Beijing, China.
関連する実験動画
Contact us if these videos are not relevant.
Contact us if these videos are not relevant.
PubMedで要約を見る
まとめ
この要約は機械生成です。この研究は,ネガティブな転送問題を克服する一般的な医療セグメンテーションモデルであるMedSegXを導入します. MedSegXは標準的および配送外の医用イメージングタスクの両方で堅実なパフォーマンスを示しています.
科学分野
- 人工知能
- 医療用イメージング
- コンピュータ・ビジョン
背景
- 視力基盤モデルは多用途の医療セグメンテーションを提供しますが,不適切な訓練前のデータのためにネガティブな転送に苦しんでいます.
- これらのモデルは,様々なオープンワールドおよびOODの医学的シナリオで限られた評価があります.
研究 の 目的
- 一般的な医療セグメンテーションモデル (MedSegX) を開発し,OODの課題に対応する.
- 訓練と評価のための包括的な医療セグメンテーションデータベース (MedSegDB) を作成する.
- MedSegXの性能を評価する
主な方法
- MedSegDBの構築 公共の129と社内の5の医療セグメンテーションデータセットからの階層的なデータベース
- オープンワールドセグメンテーションのためのコンテキストミックス・オブ・アダプター・エキスパート (ConMoAE) を利用するビジョン基盤モデルであるMedSegXの開発.
- MedSegXのインディストリビューション (ID),OOD,および現実世界の臨床セグメンテーションタスクに関する包括的な評価.
主要な成果
- MedSegXはIDの設定における様々な医療セグメンテーションタスクで最先端のパフォーマンスを達成しました.
- MedSegXは,OODと現実世界の臨床環境で強力なパフォーマンスを維持し,ゼロショットとデータ効率の良い汎用化に優れています.
- MedSegXは挑戦的な汎用化シナリオで既存の基盤モデルを上回った.
結論
- MedSegXは,ネガティブ・トランスファーとOODの制限を効果的に解決する,一般的な医療セグメンテーションの重要な進歩を表しています.
- 提案されたConMoAEアプローチは,さまざまな医療画像アプリケーションにモデルの適応性を高めます.
- MedSegDBは,医療基盤モデルの将来の研究に貴重なリソースを提供します.

