y-kitamu / gnn_dev

Graph Neural Network implementations (tensorflow)

GNN

gnn_dev's Issues

(2019) Coronary Artery Segmentation in Cardiac CT Angiography Using 3D Multi-Channel U-net

https://app.paperpile.com/my-library/Chen-et-al-2019-pGoSl2VHLRbSU3yDKfvOR0Q

CNNの入力にFrangi filteringの出力も含めて学習・推論する

(2019) Deep Vessel Segmentation By Learning Graphical Connectivity

https://app.paperpile.com/my-library/Shin-et-al-2019-pQCZoGljkBB2nbiBnbTzN6A

• CNNを学習（推論）
• 学習したCNNで血管を抽出、thinningとかでグラフ作成
• グラフをGNNの入力にしてGNN学習（推論）
• GNNの特徴量をCNNの特徴量と合わせて出力層を学習（推論）し、segmentation結果を出力

(2019) Graph Convolutional Networks for Coronary Artery Segmentation in Cardiac CT Angiography

https://app.paperpile.com/my-library/Wolterink-et-al-2019-plaYdSAQMQ9qSgJl5BI6AJA

血管の輪郭抽出にGCNを利用した論文。
cneterlineで血管グラフが作成済みなことが前提。
centerlineの各nodeに対して、輪郭の候補点をノードとして、GCNで学習する。

(2020) Learning Tree-Structured Representation for 3D Coronary Artery Segmentation

https://app.paperpile.com/my-library/Kong-et-al-2020-pCNGSsmHYQ0Kz5sb~1HALzg

ConvGRUというGRUを拡張した手法を用いてセグメンテーションを行う。
血管グラフを入力にして、セグメンテーションを出力する。

(2019) DeepCenterline: a Multi-task Fully Convolutional Network for Centerline Extraction

https://app.paperpile.com/my-library/Guo-et-al-2019-p331F9k1HARysuPbV0uHE5w

• centerlineの学習と血管の端点の学習をマルチタスクで実施。
  (血管の端点は端点をgaussianでぼかしたheatmapを学習）

• 血管の端点の推論結果から、大動脈に近い点を血管の起点、それ以外の点を終点にする

• centerlineの距離画像を重みにして、起点と終点の最短経路を構成し、グラフを作成する

• 評価は以下の指標
  i Mean centerline to centerline distance. The mean absolute distance from centerline A to centerline B is defined as the mean of the absolute distance to the nearest point on B for every point on A.
  ii Coverage percentage. A point on centerline A is covered by centerline B if the closest point on B is within half a voxel (0.2 mm).
  iii Number of missing endpoints. The number of endpoints not found by auto- mated algorithm is counted manually in patient-level as well as branch-level. iv Number of scans with wrong bifurcations. This usually happens when two branches are spatially close or even briefly joined at a certain section. The centerline could wrongly consider this brief joining as a bifurcation.
  v Average patient-level centerline length. The patient-level centerline length is computed as the sum of lengths of each centerline segment. In general, the less straight “shortcut” centerline takes, and/or the more branch endpoints are detected, the longer centerline will be.
  vi Hausdorff distance. Hausdorff distance is defined as the maximum of dis- tances from every artery segmentation mask voxel to the closest centerline point. Hausdorff distance shows how close each segmentation voxel is being covered by the extracted centerline.
  vii Overall success rate. A centerline extraction is called fully successful when an expert reviews the centerline and determines that the centerline covers all branches sufficiently, has no spurious false positive branch, no wrong bifurcation, and no obvious deviation from the center throughout all sections.