/YNet

Primary LanguagePythonMIT LicenseMIT

YNet: A Y-shape Fully Convolutional Neural Network for Box-driven Weakly Supervised Medical Image Segmentation

This is Tensorflow implementation of our paper "YNet: A Y-shape Fully Convolutional Neural Network for Box-driven Weakly Supervised Medical Image Segmentation"

Our method aiming at segment the liver lesion with the box annotation.

The architecture of our proposed YNet is shown in here img The demo result is shown in here img

Run

  1. Requirements

    • Python 2.7
    • Tensorflow 1.10.0
    • you can install the necessary packages by this command
    pip -r install requirements.txt

  2. Prepare Dataset

    • Run the ISBI2017_to_tfrecords.py to generate tfrecords files. You need to modify the path source dir and the target dir.
    python datasets/ISBI2017_to_tfrecords.py

  3. Run Training

    • train_segmentation.py (default is UNet based on VGG16)
    python -u ./train_segmentation_ISBI2017_UNet.py --train_dir=/home/give/PycharmProjects/weakly_label_segmentation/logs/ISBI2017_V2/UNet --num_gpus=1 --learning_rate=1e-4 --gpu_memory_fraction=1 --train_image_width=256 --train_image_height=256 --batch_size=4 --dataset_dir=/home/give/Documents/dataset/ISBI2017/weakly_label_segmentation_V4/Batch_2/tfrecords_V2_V2 --dataset_name=isbi2017v2 --dataset_split_name=train --max_number_of_steps=10000000 --checkpoint_path=/home/give/github/pixel_link/checkpoints/conv2_2/model.ckpt-73018 --using_moving_average=1 --decoder=upsampling
    • Explain the parameters
      • train_dir: the ckpt and tensorboard files will be saved in there.
      • train_image_width: the resized width
      • train_image_height: the resized height
      • dataset_dir: the path of tfrecord files
      • max_number_of_steps: the number of steps
      • checkpoint_path: pretrained model path

  4. Run Evaluation

    • evulate_segmentation_ISBI2017.py
    python -u ./evulate_segmentation_ISBI2017_V2.py --checkpoint_path=/home/give/PycharmProjects/weakly_label_segmentation/logs/ISBI2017_V2/1s_agumentation_weakly-upsampling-2/model.ckpt-168090 --dataset_dir=/home/give/Documents/dataset/ISBI2017/weakly_label_segmentation_V4/Batch_1 --pred_vis_dir=/home/give/Documents/dataset/ISBI2017/weakly_label_segmentation_V4/Batch_1/DLSC_0/pred_vis --pred_dir=/home/give/Documents/dataset/ISBI2017/weakly_label_segmentation_V4/Batch_1/DLSC_0/pred --recovery_img_dir=/home/give/Documents/dataset/ISBI2017/weakly_label_segmentation_V4/Batch_1/DLSC_0/recovery_img_step --recovery_feature_map_dir=/home/give/Documents/dataset/ISBI2017/weakly_label_segmentation_V4/Batch_1/DLSC_0/recovery_feature_map --decoder=upsampling --using_moving_average=1 --batch_size=2 --update_center=True --test_flag=False --num_centers_k=4 --nii_flag=False --full_annotation_flag=False