betashort / survey_paper_and_pdf

読んだ論文をまとめる
2 stars 0 forks source link

Unsupervised Segmentation of 3D Medical Images Based on Clustering and Deep Representation Learning #2

Open betashort opened 4 years ago

betashort commented 4 years ago

0. 論文情報・リンク

論文リンク:https://arxiv.org/abs/1804.03830 公開日時:Submitted on 11 Apr 2018 実装コード: Publication : Proc. SPIE 10578, Medical Imaging 2018: Biomedical Applications in Molecular, Structural, and Functional Imaging, 1057820 (12 March 2018)

1. どんなもの?

Joint unsupervised learning + kmeansという手法で、3D Medical imageをセグメントする。

2. 先行研究と比べてどこがすごいの?

NMI(Normalized Mutual Information):正規化相互情報量で、セグメンテーション精度を評価した。 従来法として、Multi-Otsu・K-meansと比較している。 従来法に比べ、全て(3つ)のデータで、一番精度が良かった。 IMG_0273

3. 技術や手法の"キモ"はどこにある?

JULE(Joint unsupervised learning) + kmeansという手法で、3D Medical imageをセグメントする。

Phase 1. JULEで特徴量の選択・抽出をする Phase 2. K-meansで、クラスラベルに割り当てる

IMG_0271 IMG_0270

4. どうやって有効だと検証した?

NMI(Normalized Mutual Information):正規化相互情報量で、セグメンテーション精度を評価した。 従来法として、Multi-Otsu・K-meansと比較している。

5. 議論はあるか?

6. 次に読むべき論文はあるか?

Joint Unsupervised Learning of Deep Representations and Image Clusters

7. 実装

8. 参考文献

betashort commented 4 years ago

Abstract

新しい3D医用画像の教師なしセグメンテーションを提案する。CNNは、画像セグメンテーションにおいて、重要な利点がある。最近のセグメンテーションは、手作業によるアノテーションデータを大量に必要とする教師あり学習に頼り切っている。

途中