Rendezvous: Attention Mechanisms for the Recognition of Surgical Action Triplets in Endoscopic Videos

C.I. Nwoye, T. Yu, C. Gonzalez, B. Seeliger, P. Mascagni, D. Mutter, J. Marescaux, and N. Padoy

This repository contains the implementation code, inference demo, and evaluation scripts.

Abstract

Out of all existing frameworks for surgical workflow analysis in endoscopic videos, action triplet recognition stands out as the only one aiming to provide truly fine-grained and comprehensive information on surgical activities. This information, presented as <instrument, verb, target> combinations, is highly challenging to be accurately identified. Triplet components can be difficult to recognize individually; in this task, it requires not only performing recognition simultaneously for all three triplet components, but also correctly establishing the data association between them.

To achieve this task, we introduce our new model, the Rendezvous (RDV), which recognizes triplets directly from surgical videos by leveraging attention at two different levels. We first introduce a new form of spatial attention to capture individual action triplet components in a scene; called Class Activation Guided Attention Mechanism (CAGAM). This technique focuses on the recognition of verbs and targets using activations resulting from instruments. To solve the association problem, our RDV model adds a new form of semantic attention inspired by Transformer networks; Multi-Head of Mixed Attention (MHMA). This technique uses several cross and self attentions to effectively capture relationships between instruments, verbs, and targets.

We also introduce CholecT50 - a dataset of 50 endoscopic videos in which every frame has been annotated with labels from 100 triplet classes. Our proposed RDV model significantly improves the triplet prediction mAP by over 9% compared to the state-of-the-art methods on this dataset.

News and Updates

[2025.09.17]: Check out our CAMMA Dataset Overlaps repository for an analysis of video overlaps across Cholec80, CholecT50, and Endoscapes to ensure fair dataset splits.
[2023.02.20]: CholecT50 can now be accessed on CAMMA website.
[2022.05.09]: TensorFlow v2 implementation code released!
[2022.05.09]: TensorFlow v1 implementation code released!
[2022.05.03]: PyTorch implementation code released!
[2022.04.12]: CholecT45: 45 videos subset of CholecT50 can now be accessed on CAMMA website.
[2022.03.22]: Paper accepted at Elsevier journal, Medical Image Analysis 2022!
[2022.04.01]: Demo code and pre-trained model released!

Model Overview

The RDV model is composed of:

Feature Extraction layer: extract high and low level features from input image from a video
Encoder: for triplet components encoding
- Weakly-Supervised Localization (WSL) Layer: for localizing the instruments
- Class Activation Guided Attention Mechanism (CAGAM): for detecting the verbs and targets leveraging attention resulting from instrument activations. (channel anad position spatial attentions are used here)
- Bottleneck layer: for collecting unfiltered features for initial scene understanding
Decoder: for triplet assocaition decoding over L successive layers
- Multi-Head of Mixed Attention (MHMA): for learning to associate instrument-verb-target using successive self- and cross-attention mechanism
- Feed-forward layer: for triplet feature refinement
Classifier: for final triplet classification

We hope this repo will help researches/engineers in the development of surgical action recognition systems. For algorithm development, we provide training data, baseline models and evaluation methods to make a level playground. For application usage, we also provide a small video demo that takes raw videos as input without any bells and whistles.

Performance

Results Table

||Components AP ||||| Association AP ||| :---:|:---:|:---:|:---: |:---:|:---:|:---:|:---:|:---:| API | APV | APT ||| APIV | APIT | APIVT | 92.0 | 60.7 | 38.3 ||| 39.4 | 36.9 | 29.9|

Video Demo

Available on Youtube.

Installation

Requirements

The model depends on the following libraries:

sklearn
PIL
Python >= 3.5
ivtmetrics
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

System Requirements:

The code has been test on Linux operating system. It runs on both CPU and GPU. Equivalence of basic OS commands such as unzip, cd, wget, etc. will be needed to run in Windows or Mac OS.

Quick Start

clone the git repository: git clone https://github.com/CAMMA-public/rendezvous.git
install all the required libraries according to chosen your framework.
download the dataset
download model's weights
train
evaluate

Docker Example

coming soon . . .

Dataset Zoo

Data Preparation

All frames are resized to 256 x 448 during training and evaluation.
Image data are mean normalized.
The dataset variants are tagged in this code as follows:
- cholect50 = CholecT50 with split used in the original paper.
- cholect50-challenge = CholecT50 with split used in the CholecTriplet challenge.
- cholect45-crossval = CholecT45 with official cross-val split (currently public released).
- cholect50-crossval = CholecT50 with official cross-val split.

Evaluation Metrics

The ivtmetrics computes AP for triplet recognition. It also support the evaluation of the recognition of the triplet components.

pip install ivtmetrics

conda install -c nwoye ivtmetrics

Usage guide is found on pypi.org.

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/CholecT50 Dataset

Simple training on CholecT50 dataset:

python run.py -t  --data_dir="/path/to/dataset" --dataset_variant=cholect50 --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 3 --batch 32 --version=1 --test_ckpt="/path/to/model-k3/weights"

Training on Custom Dataset

Adding custom datasets is quite simple, what you need to do are:

organize your annotation files in the same format as in CholecT45 dataset.
final model layers can be modified to suit your task by changing the class-size (num_tool_classes, num_verb_classes, num_target_classes, num_triplet_classes) in the argparse.

Model Zoo

N.B. Download links to models' weights will not be provided until after the CholecTriplet2022 challenge.

PyTorch

| Network | Base | Resolution | Dataset | Data split | Model Weights | ------------|-----------|------------|-----------|-------------|-------------------| | Rendezvous| ResNet-18 | Low | CholecT50 | RDV | Download | | Rendezvous| ResNet-18 | High | CholecT50 | RDV | [Download] | | Rendezvous| ResNet-18 | Low | CholecT50 | Challenge | Download | | Rendezvous| ResNet-18 | Low | CholecT50 | crossval k1 | Download | | Rendezvous| ResNet-18 |

Rendezvous

Install / Use

README