Suture detection PyTorch

This repo contains the reference implementation of suture detection model in PyTorch for the paper

Point detection through multi-instance deep heatmap regression for sutures in endoscopy

Lalith Sharan, Gabriele Romano, Julian Brand, Halvar Kelm, Matthias Karck, Raffaele De Simone, Sandy Engelhardt

Accepted, IJCARS 2021

Please see the license file for terms os use of this repo. If you find our work useful in your research please consider citing our paper:

Sharan, L., Romano, G., Brand, J. et al. Point detection through multi-instance deep heatmap regression for 
sutures in endoscopy. Int J CARS (2021). https://doi.org/10.1007/s11548-021-02523-w

Setup

A conda environment is recommended for setting up an environment for model training and prediction. There are two ways this environment can be set up:

1. Cloning conda environment (recommended)

conda env create -f suture_detection_pytorch.yml
conda activate suture_detection_pytorch

If the installation from .yml file does not work, it may be a cuda error. The solution is to either install the failed packages via pip, or use the pip requirements file here.

2. Installing requirements

conda intall --file conda_requirements.txt
conda install -c pytorch torchvision=0.7.0
pip install --r requirements.txt

Prediction of suture detection for a single image

You can predict the suture points for a single image with:

python test.py --dataroot ~/data/mkr_dataset/ --exp_dir ~/experiments/unet_baseline_fold_1/ --save_pred_points

• The command save_pred_points saves the predicted landmark co-ordinates in the resepective op folders in the ../predictions directory.
• The command save_pred_mask saves the predicted mask that is the output of the model in the resepective op folders in the ../predictions directory. The final points are extracted from this mask.

Dataset preparation

You can download the challenge dataset from the synapse platform by signing up for the AdaptOR 2021 Challenge from the Synapse platform.

• The Challenge data is present in this format: dataroot --> op_date --> video_folders --> images, point_labels
• Generate the masks with a blur function and spread by running the following script:

python generate_suture_masks.py --dataroot /path/to/data --blur_func gaussian --spread 2

• Generate the split files for the generated masks, for cross-validation by running the following script: You can predict depth for a single image with:

python generate_splits.py --splits_name mkr_dataset --num_folds 4

Training a model

Once you have prepared the dataset, you can train the model with:

python train.py --dataroot /path/to/data

• This repo is inspired by the following repos:
• CycleGAN PyTorch
• Monodepth2
• DetCycleGAN