AI-based medical image segmentation method with high versatility

• Project/Area Number: 20K20216
• Research Category: Grant-in-Aid for Early-Career Scientists
• Allocation Type: Multi-year Fund
• Review Section: Basic Section 90130:Medical systems-related
• Research Institution: Tokyo Medical and Dental University
• Principal Investigator: Sugino Takaaki 東京医科歯科大学, 生体材料工学研究所, 助教 (30830492)
• Project Period (FY): 2020-04-01 – 2022-03-31
• Project Status: Completed (Fiscal Year 2021)
• Budget Amount: ¥4,030,000 (Direct Cost: ¥3,100,000、Indirect Cost: ¥930,000); Fiscal Year 2021: ¥1,690,000 (Direct Cost: ¥1,300,000、Indirect Cost: ¥390,000); Fiscal Year 2020: ¥2,340,000 (Direct Cost: ¥1,800,000、Indirect Cost: ¥540,000)
• Keywords: セグメンテーション / 医用画像処理 / 人工知能

Outline of Research at the Start

本研究の目的は，人工知能(AI)技術の応用により，医療施設に応じて様々なデータ環境が混在する中でも汎用的に利用可能な医用画像セグメンテーション手法を開発することである．本研究では，医用画像のビッグデータ化とAIによる医用画像セグメンテーションの利活用を進める上で課題となる，教師データの量と質の確保，医用画像データの不統一性，画像処理過程のブラックボックス化の解決を図る．そのために，教師データの効率的な生成・拡張とクレンジング，医用画像データの拡張(多種化)，学習過程分析に基づくAIの最適化を可能とする技術の開発とこれらを組み合わせたセグメンテーション手法を実現する．

Outline of Final Research Achievements

The aim of this study is to develop a method for improving the versatility of medical image segmentation using artificial intelligence (AI) techniques. This study investigated the following components: a network architecture to improve the accuracy of medical image segmentation from incomplete training data with insufficient quantity and quality, a data augmentation technique to deal with the inconsistency of medical images caused by different diagnostic equipment at different medical facilities, and a loss function to stabilize the learning process and segmentation performance. As a result, this study obtained research outcomes including a network architecture and training methods that can improve the generalization performance of AI in medical image segmentation.

Academic Significance and Societal Importance of the Research Achievements

本研究の学術的意義は，量と質が不十分な教師データや利用可能な医用画像データの不統一性など，臨床環境下でのAI医用画像セグメンテーションにおける問題の改善につながる要素技術をもたらした点である．本技術は，限定的な教師データおよび医用画像からのセグメンテーション精度の向上を可能にし，教師データ作成に伴う医師の負担を軽減しながらより高精細な解剖構造の抽出・可視化が期待できる．したがって，本研究成果は，医用画像セグメンテーションの汎用的な利用とそれに伴うコンピュータ支援診断・治療の発展などにつながるものであり，大きな社会的意義がある．

Research Products

• [Journal Article] Performance improvement of weakly supervised fully convolutional networks by skip connections for brain structure segmentation (2021)
  • Author(s): Takaaki Sugino, Holger R. Roth, Masahiro Oda, Taichi Kin, Nobuhito Saito, Yoshikazu Nakajima, Kensaku Mori
  • Journal Title: Medical Physics Volume: 48 Issue: 11 Pages: 7215-7227
  • DOI: 10.1002/mp.15192
  • Peer Reviewed
• [Journal Article] Loss Weightings for Improving Imbalanced Brain Structure Segmentation Using Fully Convolutional Networks (2021)
  • Author(s): Takaaki Sugino, Toshihiro Kawase, Shinya Onogi, Taichi Kin, Nobuhito Saito, Yoshikazu Nakajima
  • Journal Title: Healthcare Volume: 9 Issue: 8 Pages: 938-938
  • DOI: 10.3390/healthcare9080938
  • Peer Reviewed / Open Access
• [Presentation] Two-stage convolutional neural network-based approach for improved segmentation of basal ganglia on MR images (2022)
  • Author(s): Takaaki Sugino, Taichi Kin, Nobuhito Saito, Yoshikazu Nakajima
  • Organizer: Computer Assisted Radiology and Surgery 2022
  • Int'l Joint Research
• [Presentation] MRI画像からの大脳基底核のAIセグメンテーション－Skip connectionによる抽出精度向上の検討－ (2021)
  • Author(s): 杉野貴明，金太一，斎藤季，川瀬利弘，小野木真哉，齊藤延人，森健策，中島義和
  • Organizer: 第30回日本コンピュータ外科学会
• [Presentation] Label cleansing and propagation for improved segmentation performance using fully convolutional networks (2020)
  • Author(s): Takaaki Sugino, Yutaro Suzuki, Taichi Kin, Nobuhito Saito, Kensaku Mori, Yoshikazu Nakajima
  • Organizer: Computer Assisted Radiology and Surgery (CARS) 2020
  • Int'l Joint Research