Publications

Scaling up self-supervised learning for improved surgical foundation models

Tim J.M. Jaspers^{1* 📧}, Ronald L.P.D. de Jong^2*, Yiping Li², Carolus H.J. Kusters¹, Franciscus H.A. Bakker⁵, Romy C. van Jaarsveld³, Gino M. Kuipers³, Richard³, Jelle P. Ruurda³, Willem M. Brinkman⁴, Josien P.W. Pluim², Peter H.N. de With¹, Marcel Breeuwer², Yasmina Al Khalil², Fons van der Sommen¹

¹ Department of Electrical Engineering, Video Coding & Architectures, Eindhoven University of Technology
² Department of Biomedical Engineering, Medical Image Analysis, Eindhoven University of Technology, Eindhoven, The Netherlands
³ Department of Surgery, University Medical Center Utrecht, Utrecht, The Netherlands
⁴ Department of Oncological Urology, University Medical Center Utrecht, Utrecht, The Netherlands
⁵ Department of Surgery, Catharina Hospital, Eindhoven, The Netherlands

^* Both authors attributed equally
(^📧) corresponding author

Medical Image Analysis
(Article)

Exploring the Effect of Dataset Diversity in Self-Supervised Learning for Surgical Computer Vision

Tim J.M. Jaspers^{1 📧}, Ronald L.P.D. de Jong², Yasmina Al Khalil², Tijn Zeelenberg ¹, Carolus H.J. Kusters¹, Franciscus H.A. Bakker⁵, Yiping Li², Romy C. van Jaarsveld³, Jelle P. Ruurda³, Willem M. Brinkman⁴, Peter H.N. de With¹, Fons van der Sommen¹,

Second Workshop on Data Engineering in Medical Imaging (DEMI) - Satellite Event MICCAI 2024
(Proceeding)

¹ Department of Electrical Engineering, Video Coding & Architectures, Eindhoven University of Technology
² Department of Biomedical Engineering, Medical Image Analysis, Eindhoven University of Technology, Eindhoven, The Netherlands
³ Department of Surgery, University Medical Center Utrecht, Utrecht, The Netherlands
⁴ Department of Oncological Urology, University Medical Center Utrecht, Utrecht, The Netherlands
⁵ Department of Surgery, Catharina Hospital, Eindhoven, The Netherlands

(^📧) corresponding author

Abstract

Foundation models have revolutionized computer vision by achieving state-of-the-art performance across diverse tasks through large-scale pretraining on extensive datasets. However, their application in surgical computer vision has been limited. This study addresses this gap by introducing SurgeNetXL, a novel surgical foundation model that sets a new benchmark in surgical computer vision. Trained on the largest reported surgical dataset to date, comprising over 4.7 million video frames, SurgeNetXL achieves consistent top-tier performance across six datasets spanning four surgical procedures and three tasks, including semantic segmentation, phase recognition, and critical view of safety (CVS) classification. Compared to the best-performing surgical foundation models, SurgeNetXL shows mean improvements of 2.4, 8.95, and 12.6% for semantic segmentation, phase recognition, and CVS classification, respectively. Additionally, SurgeNetXL outperforms the best-performing ImageNet-based variants by 14.4, 4.0, and 1.6% in the respective tasks. In addition to advancing model performance, this work provides key insights into scaling pretraining datasets, extending training durations, and optimizing model architectures specifically for surgical computer vision. These findings pave the way for improved generalizability and robustness in data-scarce scenarios, offering a comprehensive framework for future research in this domain.

Results

The following figures are from our publications, showcasing the performance of our introduced foundation model across diverse surgical tasks and procedures. These results demonstrate the model’s state-of-the-art performance on a variety of downstream tasks, reflecting its versatility and robustness in handling datasets from multiple surgical procedures.

Figure 1 and Figure 2 illustrate comparative rankings of our model against existing foundation models, highlighting its superior generalization capabilities across datasets. Figure 3 provides a t-SNE visualization, showcasing the clear cluster separation per specific dataset achieved by the model’s feature embeddings, further emphasizing its effectiveness in capturing meaningful representations.

Fig 1: Radar chart showing model ranks across datasets.

Fig 2: Blob chart representing ranking metrics for models.

Fig 3: t-SNE visualization of feature embeddings showing cluster separation across datasets.

Models

The models used in this study are based on the MetaFormer architecture. The models are trained using a self-supervised learning approach on the SurgeNetXL dataset and its variations, introduced this in the following paper. All model weights can be downloaded from the table below.

Model	Backbone	Epochs	Teacher Backbone	Full DINO checkpoint
SurgeNetXL	CaFormer	50	Download	Download
SurgeNetSmall	CaFormer	50	Download	Download
SurgeNetCholec	CaFormer	50	Download	Download
SurgeNetRAMIE	CaFormer	50	Download	Download
SurgeNetRARP	CaFormer	50	Download	Download
SurgeNetPublic	CaFormer	50	Download	Download
SurgeNet	CaFormer	50	Download	Download
SurgeNet	ConvNextv2	50	Download	Download
SurgeNet	PVTv2	50	Download	Download

Installation using Conda

We recommend using Anaconda for installation, but you are welcome to use other methods as long as the correct versions of Python and Timm are installed. If you don't have Anaconda, download it here: Anaconda Download

Create a Conda environment:

conda create --name SurgeNet python=3.8

Activate your Conda environment:

conda activate SurgeNet

Install timm using pip:

pip install timm==1.0.13

Clone our repository:

git clone https://github.com/TimJaspers0801/SurgeNet.git

Navigate to the folder:

cd SurgeNet

Load our models by running load_models.py (see explanation below).

Loading Models

The weights from the teacher network can be used to initialize either your classification or segmentation model. This can be done by running load_models.py, which contains the following code:

import torch
from metaformer import caformer_s18, MetaFormerFPN
from convnextv2 import convnextv2_tiny, ConvNextFPN
from pvtv2 import pvt_v2_b2, PVTV2FPN

urls = {
    "ImageNet1k": "https://huggingface.co/sail/dl/resolve/main/caformer/caformer_s18.pth",
    "SurgeNetXL": "https://huggingface.co/TimJaspersTue/SurgeNetModels/resolve/main/SurgeNetXL_checkpoint_epoch0050_teacher.pth?download=true",
    "SurgeNet": "https://huggingface.co/TimJaspersTue/SurgeNetModels/resolve/main/SurgeNet_checkpoint_epoch0050_teacher.pth?download=true",
    "SurgeNet-Small": "https://huggingface.co/TimJaspersTue/SurgeNetModels/resolve/main/SurgeNetSmall_checkpoint_epoch0050_teacher.pth?download=true",
    "SurgeNet-CHOLEC": "https://huggingface.co/TimJaspersTue/SurgeNetModels/resolve/main/CHOLEC_checkpoint_epoch0050_teacher.pth?download=true",
    "SurgeNet-RAMIE": "https://huggingface.co/TimJaspersTue/SurgeNetModels/resolve/main/RAMIE_checkpoint_epoch0050_teacher.pth?download=true",
    "SurgeNet-RARP": "https://huggingface.co/TimJaspersTue/SurgeNetModels/resolve/main/RARP_checkpoint_epoch0050_teacher.pth?download=true",
    "SurgeNet-Public": "https://huggingface.co/TimJaspersTue/SurgeNetModels/resolve/main/Public_checkpoint0050.pth?download=true",
    "SurgeNet-ConvNextv2": "https://huggingface.co/TimJaspersTue/SurgeNetModels/resolve/main/SurgeNet_ConvNextv2_checkpoint_epoch0050_teacher.pth?download=true",
    "SurgeNet-PVTv2": "https://huggingface.co/TimJaspersTue/SurgeNetModels/resolve/main/SurgeNet_PVTv2_checkpoint_epoch0050_teacher.pth?download=true",
}

# CAFormer model
classification_model = caformer_s18(num_classes=12, pretrained='SurgeNet', pretrained_weights=urls['SurgeNetXL'])
segmentation_model = MetaFormerFPN(num_classes=12, pretrained='SurgeNet', pretrained_weights=urls['SurgeNetXL'])

# ConvNextv2 model
classification_model = convnextv2_tiny(num_classes=12, pretrained_weights=urls['SurgeNet-ConvNextv2'])
segmentation_model = ConvNextFPN(num_classes=12, pretrained_weights=urls['SurgeNet-ConvNextv2'])

# PVTv2 model
classification_model = pvt_v2_b2(num_classes=12, pretrained_weights=urls['SurgeNet-PVTv2'])
segmentation_model = PVTV2FPN(num_classes=12, pretrained_weights=urls['SurgeNet-PVTv2'])

Note: If your want a different version of SurgeNet weights (e.g. SurgeNet-Small), you can replace the `pretrained_weights` argument with the desired url (leave the `pretrained` argument as it is).

Surgical Youtube Dataset

A key contribution of our research is the Surgical YouTube dataset, which significantly improved our foundation model’s performance. This curated dataset contains 2,074,234 frames sampled from 23 distinct surgical procedures and is available upon request at huggingface datasets. Alternatively, you can generate the dataset yourself by running create_SurgeNetXL.py together with the text file provided here. The original dataset was created by Schmidgall et al. (2024) and manually curated in our work. For video authors who prefer not to be included, an opt-out form is available. This dataset forms a large component of the SurgeNetXL dataset, which also incorporates other open-source datasets listed below.

Procedure-specific subset	Dataset	Procedure	#videos	#frames	Public
SurgeNetCholec	Cholec80 (Twinnanda et al., 2017b)	Laparoscopic Cholecystectomy	76	179,164	Yes
	HeiChole (Maier-Hein et al., 2021)	Laparoscopic Cholecystectomy	30	53,427	Yes
	hSDB-Chole (Yoon et al., 2021)	Laparoscopic Cholecystectomy	24	18,064	Yes
SurgeNetRAMIE	RAMIE-UMCU	RA Esophagectomy	28	377,287	No
SurgeNetRARP	ESAD Bawa et al., 2021	RA Esophagectomy	28	47,282	Yes
	PSI-AVA Valderrama et al., 2022	RA Prostatectomy	8	73,618	Yes
	RARP-AvL	RA Prostatectomy	8	261,516	No
Others	DSAD (Carstens et al., 2023)	RA Rectal Resection/Extirpation	32	14,623	Yes
	GLENDA (Leibetseder et al., 2020)	Gynecologic Laparoscopy	400	25,682	Yes
	LapGyn4 (Leibetseder et al., 2018)	Gynecologic Laparoscopy	500	59,616	Yes
	MultiBypass140 (Lavanchy et al., 2024)	Laparoscopic Gastric Bypass Surgery	140	749,419	Yes
	hSDB-Gastric (Yoon et al., 2021)	RA Gastrectomy	24	35,576	Yes
	SurgToolLoc2022 (Zia et al., 2023)	11 different RA porcine procedures	N/A	741,516	Yes
	YouTube ours	23 identified procedures	3,253	2,074,234	Yes

SurgeNetXL variations	Dataset	Procedure	#videos	#frames	Public
SurgeNetSmall	10% of the above (excluding YouTube)	All of the above (excluding YouTube)	>1345	263,679	Partly
SurgeNetPublic	All public datasets (excluding YouTube & private datasets)	All of the above (excluding YouTube & RA Esophagectomy)	>1238	1,997,987	Yes
SurgeNet	All of the above (excluding YouTube)	All of the above (excluding YouTube)	>1345	2,636,790	Partly
SurgeNetXL	All of the above	All of the above	>4598	4,711,024	Partly

Acknowledgements

Our implementation of the feature pyramid network is based on the pytorch segmentation models library. Pretraining on SurgeNet was performed using the code provided with the DINO publication. We have used the code of Schmidgall et al. (2024) to obtain the youtube videos, this code can be found here.

Citation

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

@article{JASPERS2026103873,
title = {Scaling up self-supervised learning for improved surgical foundation models},
journal = {Medical Image Analysis},
volume = {108},
pages = {103873},
year = {2026},
issn = {1361-8415},
doi = {https://doi.org/10.1016/j.media.2025.103873},
url = {https://www.sciencedirect.com/science/article/pii/S1361841525004190},
author = {Tim J.M. Jaspers and Ronald L.P.D. de Jong and Yiping Li and Carolus H.J. Kusters and Franciscus H.A. Bakker and Romy C. van Jaarsveld and Gino M. Kuiper and Richard van Hillegersberg and Jelle P. Ruurda and Willem M. Brinkman and Josien P.W. Pluim and Peter H.N. de With and Marcel Breeuwer and Yasmina Al Khalil and Fons van der Sommen},
keywords = {Self-supervised learning, SurgeNet, Surgical computer vision, Transfer learning},
}

@inbook{Jaspers2024,
   title={Exploring the Effect of Dataset Diversity in Self-supervised Learning for Surgical Computer Vision},
   ISBN={9783031737480},
   ISSN={1611-3349},
   url={http://dx.doi.org/10.1007/978-3-031-73748-0_5},
   DOI={10.1007/978-3-031-73748-0_5},
   booktitle={Data Engineering in Medical Imaging},
   publisher={Springer Nature Switzerland},
   author={Jaspers, Tim J. M. and de Jong, Ronald L. P. D. and Al Khalil, Yasmina and Zeelenberg, Tijn and Kusters, Carolus H. J. and Li, Yiping and van Jaarsveld, Romy C. and Bakker, Franciscus H. A. and Ruurda, Jelle P. and Brinkman, Willem M. and De With, Peter H. N. and van der Sommen, Fons},
   year={2024},
   month=oct, pages={43–53} }