| 講演名 | 2022-07-08 [Short Paper] Multi-phase CT Image Segmentation with Single-Phase Annotation Using Adversarial Unsupervised Domain Adaptation Swathi Ananda(Swathi), Yutaro Iwamoto(Yutaro), Xianhua HAN(Xianhua), Lanfen Lin(Lanfen), Hongjie Hu(Hongjie), Yen-Wei Chen(Yen-Wei), |
|---|---|
| PDFダウンロードページ |  PDFダウンロードページへ |
| 抄録(和) | Multi-phase computed tomography (CT) images are widely used for the diagnosis of liver disease, since different phase has different contrast enhancement (i.e., different domain), the multi-phase CT images should be annotated for all phase images for liver or tumor segmentation, which is a time-consuming and labor-expensive task. To lower the cost of manual annotation and domain shift problem, we propose an adversarial unsupervised domain adaptation (UDA) method for liver segmentation of multi-phase CT images with only single-phase annotation. The framework consists of two modules: a generator and a discriminator. We have employed U-Net as a generator as it is designed for medical image segmentation. We first use the annotated source images to train the generator only. Then we use the adversarial learning to train both generator and discriminator to minimize the difference between the source heatmap and target heatmap (domain shift). The refined generator is used for multi-phase CT image segmentation. To perform liver segmentation, initially, we have conducted experiments within each phase of our internal MPCT-FLL dataset (i.e., by utilizing the PV phase as the source and the ART phase as the target). Further, we use the publicly available LiTS dataset which consists of annotated PV phase images as the source domain, and each phase of our internal MPCT-FLL dataset i.e., (PV, ART, NC phase) as a target domain. The experimental results of this work suggest consistent and comparable improvement in the performance of our liver tumor segmentation over the previously reported state-of-the-art methods. |
| 抄録(英) | Multi-phase computed tomography (CT) images are widely used for the diagnosis of liver disease, since different phase has different contrast enhancement (i.e., different domain), the multi-phase CT images should be annotated for all phase images for liver or tumor segmentation, which is a time-consuming and labor-expensive task. To lower the cost of manual annotation and domain shift problem, we propose an adversarial unsupervised domain adaptation (UDA) method for liver segmentation of multi-phase CT images with only single-phase annotation. The framework consists of two modules: a generator and a discriminator. We have employed U-Net as a generator as it is designed for medical image segmentation. We first use the annotated source images to train the generator only. Then we use the adversarial learning to train both generator and discriminator to minimize the difference between the source heatmap and target heatmap (domain shift). The refined generator is used for multi-phase CT image segmentation. To perform liver segmentation, initially, we have conducted experiments within each phase of our internal MPCT-FLL dataset (i.e., by utilizing the PV phase as the source and the ART phase as the target). Further, we use the publicly available LiTS dataset which consists of annotated PV phase images as the source domain, and each phase of our internal MPCT-FLL dataset i.e., (PV, ART, NC phase) as a target domain. The experimental results of this work suggest consistent and comparable improvement in the performance of our liver tumor segmentation over the previously reported state-of-the-art methods. |
| キーワード(和) | Multi-phase CT image / domain adaptation / adversarial learrning |
| キーワード(英) | Multi-phase CT image / domain adaptation / adversarial learrning |
| 資料番号 | MI2022-38 |
| 発行日 | 2022-07-01 (MI) |
| 研究会情報 | |
| 研究会 | MI |
|---|---|
| 開催期間 | 2022/7/8(から2日開催) |
| 開催地（和） | 小樽商工会議所・小樽経済センターホール（４階） |
| 開催地（英） | |
| テーマ（和） | 医用画像処理・認識一般 |
| テーマ（英） | Medical imaging, recoginition, etc. |
| 委員長氏名（和） | 本谷 秀堅(名工大) |
| 委員長氏名（英） | Hidekata Hontani(Nagoya Inst. of Tech.) |
| 副委員長氏名（和） | 羽石 秀昭(千葉大) / 北坂 孝幸(愛知工大) |
| 副委員長氏名（英） | Hideaki Haneishi(Chiba Univ.) / Takayuki Kitasaka(Aichi Inst. of Tech.) |
| 幹事氏名（和） | 平野 靖(山口大) / 原口 亮(兵庫県立大) |
| 幹事氏名（英） | Yasushi Hirano(Yamaguchi Univ.) / Ryo Haraguchi(Univ. of Hyogo) |
| 幹事補佐氏名（和） | 原 武史(岐阜大) / 大竹 義人(奈良先端大) |
| 幹事補佐氏名（英） | Takeshi Hara(Gifu Univ.) / Yoshito Otake(NAIST) |
| 講演論文情報詳細 | |
| 申込み研究会 | Technical Committee on Medical Imaging |
|---|---|
| 本文の言語 | ENG |
| タイトル（和） | |
| サブタイトル（和） | |
| タイトル（英） | [Short Paper] Multi-phase CT Image Segmentation with Single-Phase Annotation Using Adversarial Unsupervised Domain Adaptation |
| サブタイトル（和） | |
| キーワード(1)（和/英） | Multi-phase CT image / Multi-phase CT image |
| キーワード(2)（和/英） | domain adaptation / domain adaptation |
| キーワード(3)（和/英） | adversarial learrning / adversarial learrning |
| 第 1 著者 氏名（和/英） | Swathi Ananda / Swathi Ananda |
| 第 1 著者 所属（和/英） | Ritsumeikan University(略称：Swathi) Ritsumeikan University(略称：Swathi) |
| 第 2 著者 氏名（和/英） | Yutaro Iwamoto / Yutaro Iwamoto |
| 第 2 著者 所属（和/英） | Ritsumeikan University, Shiga, Japan(略称：Yutaro) Ritsumeikan University, Shiga, Japan(略称：Yutaro) |
| 第 3 著者 氏名（和/英） | Xianhua HAN / Xianhua HAN |
| 第 3 著者 所属（和/英） | Yamaguchi University, Japan(略称：Xianhua) Yamaguchi University, Japan(略称：Xianhua) |
| 第 4 著者 氏名（和/英） | Lanfen Lin / Lanfen Lin |
| 第 4 著者 所属（和/英） | Zhejiang University, Hangzhou, China(略称：Lanfen) Zhejiang University, Hangzhou, China(略称：Lanfen) |
| 第 5 著者 氏名（和/英） | Hongjie Hu / Hongjie Hu |
| 第 5 著者 所属（和/英） | Zhejiang University, Hangzhou, China(略称：Hongjie) Zhejiang University, Hangzhou, China(略称：Hongjie) |
| 第 6 著者 氏名（和/英） | Yen-Wei Chen / Yen-Wei Chen |
| 第 6 著者 所属（和/英） | Ritsumeikan University, Shiga, Japan(略称：Yen-Wei) Ritsumeikan University, Shiga, Japan(略称：Yen-Wei) |
| 発表年月日 | 2022-07-08 |
| 資料番号 | MI2022-38 |
| 巻番号（vol） | vol.122 |
| 号番号（no） | MI-98 |
| ページ範囲 | pp.24-25(MI), |
| ページ数 | 2 |
| 発行日 | 2022-07-01 (MI) |