lzhLab/vesselSeg

vesselSeg

This study presents an imbalanced bidirectional-scaling enhanced attention model for liver vessel segmentation, of which the shallow down-scaling module enlarges the receptive field and suppresses intensive pixel-level noise, the deep up-scaling module is a super-resolution architecture aiming at zooming in vessel details, and the attention module is to capture structural connections.

Usage

Parameters

• num_workers: int
  Number of workers. Used to set the number of threads to load data.
• ckpt: str
  Weight path. Used to set the dir path to save model weight.
• w: str
  The path of model wight to test or reload.
• heads: int
  Number of heads in Multi-head Attention layer.
• mlp_dim: int.
  Dimension of the MLP (FeedForward) layer.
• channels: int, default 3.
  Number of image's channels.
• dim: int.
  Last dimension of output tensor after linear transformation nn.Linear(..., dim).
• dropout: float between [0, 1], default 0.
  Dropout rate.
• emb_dropout: float between [0, 1], default 0.
  Embedding dropout rate.
• patch_h and patch_w:int
  The patches size.
• dataset_path: str
  Used to set the relative path of training and validation set.
• batch_size: int
  Batch size.
• max_epoch: int
  The maximum number of epoch for the current training.
• lr: float
  learning rate. Used to set the initial learning rate of the model.