MF-TAPNet

---

Useful Links

MICCAI 2017 Endoscopic vision challenge

Dataset and instructions

Download dataset and report summary

As required, you need to register an account and get the permission to download the dataset.

Instructions

To train and evaluate the MF-TAPNet model, you may follow the instructions.

Dependencies

- Python 3.6
- pytorch 0.4.1+
- pytorch-ignite 0.2.0+
- tensorboardX
- albumentations
- opencv-python
- cupy (please check your CUDA version before install)
- tqdm

Folder Structure

The structure of this project will be arranged as follows:

(root folder)
├── data
|  ├── train
|  |  ├── instrument_dataset_1
|  |  |  ├── left_frames
|  |  |  ├── right_frames
|  |  |  ├── ......
|  |  ├── instrument_dataset_2
|  |  |  ├── left_frames
|  |  |  ├── right_frames
|  |  |  ├── ......
|  ├── cropped_train
├── src
├── pretrained_model
|  ├── network-css.pytorch
├── ......

Download prerequisites

1. Assume the working directory is $ROOT_DIR/. Download the train dataset (2 zips including instrument_dataset 1 to 8). Unzip instrument_dataset_X into data/train following the above structure. The train dataset should be arranged as $ROOT_DIR/data/train/instrument_dataset_X/.... Note that it's almost the same for the test dataset, and should be arranged as $ROOT_DIR/data/test/instrument_dataset_X/....

$ git clone https://github.com/keyuncheng/MF-TAPNet.git
$ cd MF-TAPNet
$ mkdir data/
$ mkdir data/train
$ unzip instrument_1_4_training.zip -d data/train
$ unzip instrument_5_8_training.zip -d data/train

2. Download UnFlow pytorch pretrained model for optical flow estimation, then move it to $ROOT_DIR/pretrained_model/.

$ wget --timestamping http://content.sniklaus.com/github/pytorch-unflow/network-css.pytorch
$ mkdir pretrained_model
$ mv network-css.pytorch pretrained_model

3. Switch to source code folder

$ cd src/

Preprocess data

1. preprocess training dataset (images and masks).

$ python preprocess_data.py

for test dataset:

$ python preprocess_data.py --data_dir ../data/test/ --cropped_data_dir ../data/cropped_test --mode test

2. Estimate optical flow for image pairs

We use pretrained UnFlow to estimate optical flow for consecutive image pairs in each surgical video.

$ python gen_optflow.py

Note: This step is tricky because the UnFlow model are pretrained using datasets with optical flow ground truths (KITTI, 1280 * 384). However, we are trying to estimate the optical flow using surgical videos frames in different sizes. In addition, we cannot train-from-scratch/finetune the UnFlow model in a supervised way without dense optical flow as ground truths. For more accurate optical flow estimation, we are trying other methods (unsupervised fine-tuning using surgical videos).

Train the model

Arguments for model training in train.sh are in default settings. You may try other models from /models/plane_model.py or /models/tap_model.py by modifying the p--model] argument.

$ sh train.sh

After training all the folds, the program will show validation statistics (mean IoU, mean Dice)

Semi-supervised learning

Set arguments for model training in train.sh with "--semi true". A sample in train.sh for semi-supervised learning is provided for reference.

Issues

This repository is still updating in progress. We will keep the code updated for any problems encountered.