[AAAI 2025] Surgical Workflow Recognition and Blocking Effectiveness Detection in Laparoscopic Liver Resections with Pringle Maneuver

Abstract

Pringle maneuver (PM) in laparoscopic liver resection aims to reduce blood loss and provide a clear surgical view by intermittently blocking blood inflow of the liver, whereas prolonged PM may cause ischemic injury. To comprehensively monitor this surgical procedure and provide timely warnings of ineffective and prolonged blocking, we suggest two complementary AI-assisted surgical monitoring tasks: workflow recognition and blocking effectiveness detection in liver resections. The former presents challenges in real-time capturing of short-term PM, while the latter involves the intraoperative discrimination of long-term liver ischemia states. To address these challenges, we meticulously collect a novel dataset, called PmLR50, consisting of 25,037 video frames covering various surgical phases from 50 laparoscopic liver resection procedures. Additionally, we develop an online baseline for PmLR50, termed PmNet. This model embraces Masked Temporal Encoding (MTE) and Compressed Sequence Modeling (CSM) for efficient short-term and long-term temporal information modeling, and embeds Contrastive Prototype Separation (CPS) to enhance action discrimination between similar intraoperative operations. Experimental results demonstrate that PmNet outperforms existing state-of-the-art surgical workflow recognition methods on the PmLR50 benchmark. Our research offers potential clinical applications for the laparoscopic liver surgery community.

🔥🔥🔥 News!!

• Dec 12, 2024: 🤗 Our work has been accepted by AAAI 2025! Congratulations!
• Feb 21, 2025: 🚀 Code and dataset have been released! Dataset Link

PmLR50 Dataset and PmNet

Installation

• Environment: CUDA 12.5 / Python 3.8
• Device: Two NVIDIA GeForce RTX 4090 GPUs
• Create a virtual environment

git clone https://github.com/RascalGdd/PmNet.git
cd PmNet
conda env create -f PmNet.yml
conda activate Pmnet

Prepare your data

Download processed data from PmLR50; The final structure of datasets should be as following:

data/
    └──PmLR50/
        └──frames/
            └──01
                ├──00000000.jpg
                ├──00000001.jpg
                └──...
            ├──...    
            └──50
        └──phase_annotations/
            └──01.txt
            ├──02.txt
            ├──...
            └──50.txt
        └──blocking_annotations/
            └──01.txt
            ├──02.txt
            ├──...
            └──50.txt
        └──bbox_annotations/
            └──01.json
            ├──02.json
            ├──...
            └──50.json

Then, process the data with generate_labels_pmlr.py to generate labels for training and testing.

Training

We provide the script for training train.sh and testing test.sh.

run the following code for training

sh train.sh

and run the following code for testing

sh test.sh

The checkpoint of our model is provided here.

More Configurations

We list some more useful configurations for easy usage:

Argument	Default	Description
--nproc_per_node	2	Number of nodes used for training and testing
--batch_size	8	The batch size for training and inference
--epochs	50	The max epoch for training
--save_ckpt_freq	10	The frequency for saving checkpoints during training
--nb_classes	5	The number of classes for surgical workflows
--data_strategy	online	Online/offline mode
--num_frames	20	The number of consecutive frames used
--sampling_rate	8	The sampling interval for comsecutive frames
--enable_deepspeed	True	Use deepspeed to accelerate
--dist_eval	False	Use distributed evaluation to accelerate
--load_ckpt	--	Load a given checkpoint for testing

Acknowledgements

Huge thanks to the authors of following open-source projects:

• TMRNet
• Surgformer
• TimeSformer

License & Citation

If you find our work useful in your research, please consider citing our paper.