/EndoVis2018

MICCAI challenge for EndoVis2018

Primary LanguageJupyter Notebook

EndoVis2018

MICCAI challenge for EndoVis2018. The challenge focuses on surgical scene segmentation.

Final Report

See pdf for more details.

Demo video

Result

Mini-Data (20% uniform subsampling)

Network 0 1 2 3 4 5 6 7 8 9 10 11 Mean
UNet 0.87 0.86 0.73 0.76 0.82 0.84 0.68 0.85 0.00 0.66 0.88 0.58 0.71
AlbuNet 0.92 0.91 0.80 0.79 0.90 0.90 0.68 0.78 0.00 0.76 0.91 0.71 0.76
AlbuNet+SuperLabel 0.93 0.93 0.82 0.80 0.91 0.90 0.62 0.86 0.00 0.78 0.92 0.77 0.77
DeepLabV3+ 0.91 0.93 0.81 0.82 0.94 0.87 0.51 0.60 0.00 0.76 0.92 0.73 0.73
DeepLabv3+SuperLabel 0.93 0.93 0.83 0.79 0.91 0.90 0.64 0.85 0.00 0.79 0.92 0.82 0.78
DeepLabV3+Aug 0.90 0.94 0.80 0.84 0.94 0.84 0.53 0.68 0.00 0.59 0.81 0.81 0.72
DeepLabv3+SuperLabel+Aug 0.94 0.93 0.83 0.81 0.92 0.92 0.64 0.84 0.00 0.81 0.94 0.83 0.78

Sequence sample

Network 0 1 2 3 4 5 6 7 8 9 10 11 Mean
UNet 0.66 0.87 0.76 0.77 0.41 0.22 0.35 0.22 0.00 0.09 0.53 0.00 0.41
AlbuNet 0.69 0.90 0.76 0.78 0.51 0.29 0.38 0.15 0.00 0.23 0.59 0.01 0.44
AlbuNet+SuperLabel 0.75 0.94 0.79 0.84 0.60 0.43 0.43 0.45 0.00 0.45 0.62 0.00 0.53
DeepLabV3+ 0.74 0.89 0.76 0.80 0.65 0.29 0.30 0.40 0.00 0.06 0.56 0.00 0.45
DeepLabv3+SuperLabel 0.74 0.92 0.78 0.83 0.64 0.33 0.33 0.39 0.00 0.20 0.59 0.00 0.48

Random sample

Network 0 1 2 3 4 5 6 7 8 9 10 11 Mean
AlbuNet+SuperLabel 0.96 0.96 0.9 0.88 0.96 0.95 0.72 0.9 0 0.83 0.96 0.87 0.82
DeepLabV3+ 0.96 0.95 0.89 0.87 0.96 0.96 0.69 0.9 0.37 0.84 0.97 0.86 0.85
DeepLabv3+SuperLabel 0.97 0.96 0.89 0.87 0.96 0.96 0.7 0.9 0.38 0.82 0.96 0.89 0.86

Visual comparison of with/without superlabel

see Comparison.ipynb

Instruction for Tensorboardx

pip install tensorboardX
pip install tensorflow

start tensorboard by tensorboard --logdir=<dir_to_store_log_file>

Example Notebook

  1. UNet
  2. AlbuNet
  3. DeepLabV3+
  4. AlbuNet SuperLabel
  5. DeepLabV3+ SuperLabel

Architecture

  1. UNet
  2. AlbuNet
  3. DeepLabV3+
  4. AlbuNet SuperLabel
  5. DeepLabV3+ SuperLabel

Components

  1. model_training.py
  2. dataset.py
  3. dice_loss.py
  4. etc.