Keep Your Eye on the Best: Contrastive Regression Transformer for Skill Assessment in Robotic Surgery

Authors: Dimitrios Anastasiou, Yueming Jin, Danail Stoyanov, and Evangelos Mazomenos

Code for Keep Your Eye on the Best: Contrastive Regression Transformer for Skill Assessment in Robotic Surgery, published in IEEE Robotics and Automation Letters (RA-L), February, 2023.

To be presented in IROS 2023, Detroit, Michigan, USA.

Introduction

We propose a novel video-based, contrastive regression architecture, Contra-Sformer, for automated surgical skill assessment in robot-assisted surgery. The proposed framework is structured to capture the differences in the surgical performance, between a test video and a reference video which represents optimal surgical execution. A feature extractor combining a spatial component (ResNet-18), supervised on frame-level with gesture labels, and a temporal component (TCN), generates spatio-temporal feature matrices of the test and reference videos. These are then fed into an actionaware Transformer with multi-head attention (A-Transformer) that produces inter-video contrastive features at frame level, representative of the skill similarity/deviation between the two videos. Validated on the JIGSAWS dataset, Contra- Sformer achieves competitive performance (Spearman Correlation 0.65 - 0.89), with a normalized mean absolute error between 5.8% - 13.4% on all tasks and across validation setups.

When optimized, Contra-Sformer generates features that faithfully represent the similarity/deviation between the two executions and encode information indicative of suboptimal execution/errors, without requiring explicit error annotations. This is validated against manual error annotations from Hutchinson et. al, and can be exploited for providing targeted feedback and real-time assessment to trainees. Example video link.

Packages

To set up a conda environment using the provided env.yaml file, simply run:

conda env create -f env.yml

Data preparation

We provide the extracted features from the ResNet-18 (resnet18_ftrs.zip). To request them, please send an email to [email protected].

Also, follow this link and download splits.zip. This file contains .csv files specifying the train and test samples along with their GRS labels for each fold of the cross-validation schemes and tasks.

Place resnet18_ftrs.zip and splits.zip in the root directory and unzip them. Then, your file structure should look like this:

.
├── resnet18_ftrs               # directory containing the extracted features from the ResNet-18 for every task/cross-val scheme/fold
│   ├── knot_tying
|   |   ├── loso
|   |   |    ├── 1out
|   |   |        ├── KT_B001.mat # features stored in .mat format
|   |   |        .
|   |   |    .   .
|   |   |    .   .
|   |   |    .   └── KT_I005.mat
|   |   |     
|   |   |     
|   |   |    └── 5out
|   |   |
|   |   ├── louo
|   |   └── 4fold
|   |
│   ├── needle_passing         # same structure as above
│   └── suturing               # same structure as above
|
├── splits
│   ├── knot_tying
|   |   ├── loso
|   |   |    ├── 1out
|   |   |        ├── train.csv
|   |   |    .   └── val.csv
|   |   |    .  
|   |   |    .  
|   |   |    
|   |   |    └── 5out
|   |   ├── louo
|   |   ├── 4fold
|   |   └── ref.csv            # reference video along with its label
|   |
│   ├── needle_passing         # same structure as above
│   └── suturing               # same structure as above
|
.
.
.
└── utils.py

Models

We also provide the saved models in .pth format (download here).

Run the code

To train and evaluate the model, simply run experiment.py and specify the task (knot_tying/needle_passing/suturing) and the cross-validation scheme (loso/louo/4fold).

python experiment.py suturing loso

To cite our work please use:

@ARTICLE{10037203,
  author={Anastasiou, Dimitrios and Jin, Yueming and Stoyanov, Danail and Mazomenos, Evangelos},
  journal={IEEE Robotics and Automation Letters}, 
  title={Keep Your Eye on the Best: Contrastive Regression Transformer for Skill Assessment in Robotic Surgery}, 
  year={2023},
  volume={8},
  number={3},
  pages={1755-1762},
  doi={10.1109/LRA.2023.3242466}}