Semi-supervised-surgical-tool-detection

This repository contains code for our paper titled "A semi-supervised teacher-student framework for surgical tool detection and localization"

in this paper we introduce a semi-supervised learning (SSL) framework in surgical tool detection paradigm which aims to mitigate the scarcity of training data and the data imbalance through a knowledge distillation approach. In our framework, we train a model with labeled data which initialises the teacher-student joint learning, where the student is trained on teacher-generated pseudo labels from unlabeled data. We propose a multi-class distance with a margin based classification loss function in the region-of-interest head of the detector to effectively segregate foreground classes from background region

Results

Dataset Download

m2cai16-tool locations dataset can be downloaded here Dataset annotations are in VOC format. However, this work uses coco format. All the required code files for voc to coco conversion can be found in data folder.

The folder structure to keep images and labels is given as follows.

 |--Tool_detection
    |-- Datasets
        |-- coco
            |-- train2017
            |-- val2017
            |-- test2017
        |-- annotations
            |-- instance_train2017.json
            |-- instances_val2017.json
            |-- instances_test2017.json

Installation

Build Environment

# create conda env
conda create -n tool python=3.6 
# activate the enviorment 
conda activate tool
# install PyTorch >=1.5 with GPU 
conda install pytorch torchvision -c pytorch 
# install detectron2 
https://github.com/facebookresearch/detectron2/blob/main/INSTALL.md   

Training

• To train the network on 1% labeled data setting, use following.
  

  CUDA_VISIBLE_DEVICES =1,2 python train_net.py \
                                    --num-gpus 2 \
                                    --config configs/coco_supervision/faster_rcnn_R_50_FPN_sup1_run1.yaml \
                                    SOLVER.IMG_PER_BATCH_LABEL 4 SOLVER.IMG_PER_BATCH_UNLABEL 4

• Just change the config file to train on different percentages of labeled set.
  

Evaluation

• To evaluate the model, use the following.
  

CUDA_VISIBLE_DEVICES =1,2 python train_net.py \
                                  --eval-only \
                                  --num-gpus 2 \
                                  --config configs/coco_supervision/faster_rcnn_R_50_FPN_sup1_run1.yaml\
                                  SOLVER.IMG_PER_BATCH_LABEL 4 SOLVER.IMG_PER_BATCH_UNLABEL 4 \
                                  MODEL.WEIGHTS path_to_checkpoint/checkpoint \
                                  

Resume Training

• To resume model training, use the following.
  

CUDA_VISIBLE_DEVICES =1,2 python train_net.py \
                                  --resume \
                                  --num-gpus 2 \
                                  --config configs/coco_supervision/faster_rcnn_R_50_FPN_sup1_run1.yaml\
                                  SOLVER.IMG_PER_BATCH_LABEL 4 SOLVER.IMG_PER_BATCH_UNLABEL 4 \
                                  MODEL.WEIGHTS path_to_checkpoint/checkpoint \
                                  

Model weights

Backbone	Supervision	Batch Size	mAP_50:95	Model Weights
ResNet50-FPN	1%	4 labeled + 4 unlabeled	20.094	link
ResNet50-FPN	2%	4 labeled + 4 unlabeled	32.311	link
ResNet50-FPN	5%	4 labeled + 4 unlabeled	42.392	link
ResNet50-FPN	10%	4 labeled + 4 unlabeled	46.886	link

Paired t-test

We performed paired t-test to determine how well proposed model performs with respect to the state-of-the-art. Code with details can be found here

Citing semi-supervised tool detection

@misc{ali2022semisupervised,
      title={A semi-supervised Teacher-Student framework for surgical tool detection and localization}, 
      author={Mansoor Ali and Gilberto Ochoa-Ruiz and Sharib Ali},
      year={2022},
      eprint={2208.09926},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}
                                  

FAQ

If anyone wants to reproduce the code and encounters a problem or wants to give a suggestion, feel free to contact me at my email [a01753093@tec.mx]