GUARDIAN: Guarding Against Uncertainty and Adversarial Risks in Robot-Assisted Surgeries (MICCAI'24 - UNSURE)

GUARDIAN: Guarding Against Uncertainty and Adversarial Risks in Robot-Assisted Surgeries

Ufaq Khan, Umair Nawaz, Tooba Tehreem Sheikh, Asif Hanif and Mohammad Yaqub

---

GUARDIAN GUARDIAN is a three-step approach to enhance model robustness and predictive accuracy. It begins with transforming clean samples into perturbed ones via adversarial attacks, posing it as a maximization problem for the model, $\mathcal{M}_\theta$. The process proceeds with two training phases: enhancing tool detection through cross-task transferability of adversarial examples and refining triplet recognition with live adversarial training. Here, $\mathcal{L_YOLO}$ denotes the combination of classification and bounding-box loss. The final step applies conformal prediction post-training, evaluating prediction reliability, with the dotted line indicating gradient updates.

---

Guardian in Action The poisoned model $\mathcal{M}_\theta$ behaves normally on clean images $\mathrm{x}$ , predicting the correct label (highlighted in green). However, when trigger noise $\delta$ is added to the image, the model instead predicts the wrong label (highlighted in red). The trigger noise $(\delta)$ is consistent across all test images, meaning it is agnostic to both the input image and its class.

Updates 🚀

• Oct 03, 2024 : Final code and models will be released soon
• July 15, 2024 : Accepted in UNSURE - MICCAI 2024   🎊 🎉

Table of Contents

• Installation
• Models
• Datasets
• Code Structure
• Run Experiments
• Results
• Citation
• Contact
• Acknowledgement

Installation 🔧

The model depends on the following libraries:

1. sklearn
2. PIL
3. Python >= 3.5
4. ivtmetrics
5. Developer's framework:
   1. For Tensorflow version 1:
      • TF >= 1.10
   2. For Tensorflow version 2:
      • TF >= 2.1
   3. For PyTorch version:
      • Pyorch >= 1.10.1
      • TorchVision >= 0.11

Steps to install dependencies

1. Create conda environment

conda create --name aiproject python=3.8
conda activate aiproject

2. Install PyTorch and other dependencies

pip install -r requirements.txt

Models 🔳

We have shown the efficacy of Guardian on two type of models. One being the recongition model and the second for task of object detection. The detection model is used to validate the cross-task transferability of our attacks.:

Rendezvous   YOLOv8

Dataset

The dataset folders for Synapse should be organized as follows:

This folder includes:

• CholecT45 dataset:
  • data: 45 cholecystectomy videos
  • triplet: triplet annotations on 45 videos
  • instrument: tool annotations on 45 videos
  • verb: action annotations on 45 videos
  • target: target annotations on 45 videos
  • dict: id-to-name mapping files
  • a LICENCE file
  • a README file

Expand this to visualize the dataset directory structure.

  ──CholecT45
      ├───data
      │   ├───VID01
      │   │   ├───000000.png
      │   │   ├───000001.png
      │   │   ├───000002.png
      │   │   ├───
      │   │   └───N.png
      │   ├───VID02
      │   │   ├───000000.png
      │   │   ├───000001.png
      │   │   ├───000002.png
      │   │   ├───
      │   │   └───N.png
      │   ├───
      │   ├───
      │   ├───
      │   |
      │   └───VIDN
      │       ├───000000.png
      │       ├───000001.png
      │       ├───000002.png
      │       ├───
      │       └───N.png
      |
      ├───triplet
      │   ├───VID01.txt
      │   ├───VID02.txt
      │   ├───
      │   └───VIDNN.txt
      |
      ├───instrument
      │   ├───VID01.txt
      │   ├───VID02.txt
      │   ├───
      │   └───VIDNN.txt
      |
      ├───verb
      │   ├───VID01.txt
      │   ├───VID02.txt
      │   ├───
      │   └───VIDNN.txt
      |
      ├───target
      │   ├───VID01.txt
      │   ├───VID02.txt
      │   ├───
      │   └───VIDNN.txt
      |
      ├───dict
      │   ├───triplet.txt
      │   ├───instrument.txt
      │   ├───verb.txt
      │   ├───target.txt
      │   └───maps.txt
      |
      ├───LICENSE
      └───README.md

Dataset	Link
CholecT45	Download
m2cai16-tool-locations	Download

Running the Model

The code can be run in a trianing mode (-t) or testing mode (-e) or both (-t -e) if you want to evaluate at the end of training :

Training on CholecT45 Dataset

Simple training on CholecT45 dataset:

python run.py -t  --data_dir="/path/to/dataset" --dataset_variant=cholect45-crossval --version=1

You can include more details such as epoch, batch size, cross-validation and evaluation fold, weight initialization, learning rates for all subtasks, etc.:

python3 run.py -t -e  --data_dir="/path/to/dataset" --dataset_variant=cholect45-crossval --kfold=1 --epochs=180 --batch=64 --version=2 -l 1e-2 1e-3 1e-4 --pretrain_dir='path/to/imagenet/weights'

All the flags can been seen in the run.py file. The experimental setup of the published model is contained in the paper.

Testing

python3 run.py -e --data_dir="/path/to/dataset" --dataset_variant=cholect45-crossval --kfold 1 --batch 32 --version=1 --test_ckpt="/path/to/model-k3/weights"

Dataset	Model	Link
CholecT45 Cross-Val	Rendezvous $(\mathcal{M})$	Download

Citation ⭐

If you find our work useful, please consider giving a star ⭐ and citation.

@inproceedings{khanguardian,
  title={GUARDIAN: Guarding Against Uncertainty and Adversarial Risks in Robot-Assisted Surgeries},
  author={Khan, Ufaq and Nawaz, Umair and Sheikh, Tooba and Hanif, Asif and Yaqub, Mohammad},
  booktitle={Uncertainty for Safe Utilization of Machine Learning in Medical Imaging-6th International Workshop}
  year={2024}
}

Contact 📫

Should you have any question, please create an issue on this repository or contact us at [email protected]

---