• Demonstration of effectiveness of SSL for surgical computer vision using the largest dataset reported to date. • Strong generalization and robust evaluation are shown across six surgical datasets, four procedures, and three tasks, outperforming current SOTA foundation models. • Providing insights into large-scale SSL for surgical computer vision in terms of scaling, pretraining time, dataset composition, and model architecture. • Release of the models and a curated dataset of 2.1 million surgical video frames, establishing a critical resource for advancing surgical foundation model training Foundation models have revolutionized computer vision by achieving vastly superior 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, surgical phase recognition, and critical view of safety (CVS) classification. Compared with the best-performing surgical foundation model, SurgeNetXL shows mean improvements of 4.0%, 8.9%, and 11.4% for semantic segmentation, phase recognition, and CVS classification, respectively. Additionally, SurgeNetXL outperforms ImageNet1k by 16.1%, 8.0%, and 4.3% for the respective tasks. In addition to advancing model performance, this study 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 generalization and robustness in data-scarce scenarios, offering a comprehensive framework for future research in this domain. All models and a subset of the SurgeNetXL dataset, including over 2 million video frames, are publicly available at: https://github.com/TimJaspers0801/SurgeNet .

Gastrointestinal endoscopic image analysis presents significant challenges, such as considerable variations in quality due to the challenging in-body imaging environment, the often-subtle nature of abnormalities with low interobserver agreement, and the need for real-time processing. These challenges pose strong requirements on the performance, generalization, robustness and complexity of deep learning-based techniques in such safety-critical applications. While Convolutional Neural Networks (CNNs) have been the go-to architecture for endoscopic image analysis, recent successes of the Transformer architecture in computer vision raise the possibility to update this conclusion. To this end, we evaluate and compare clinically relevant performance, generalization and robustness of state-of-the-art CNNs and Transformers for neoplasia detection in Barrett's esophagus. We have trained and validated several top-performing CNNs and Transformers on a total of 10,208 images (2,079 patients), and tested on a total of 7,118 images (998 patients) across multiple test sets, including a high-quality test set, two internal and two external generalization test sets, and a robustness test set. Furthermore, to expand the scope of the study, we have conducted the performance and robustness comparisons for colonic polyp segmentation (Kvasir-SEG) and angiodysplasia detection (Giana). The results obtained for featured models across a wide range of training set sizes demonstrate that Transformers achieve comparable performance as CNNs on various applications, show comparable or slightly improved generalization capabilities and offer equally strong resilience and robustness against common image corruptions and perturbations. These findings confirm the viability of the Transformer architecture, particularly suited to the dynamic nature of endoscopic video analysis, characterized by fluctuating image quality, appearance and equipment configurations in transition from hospital to hospital. The code is made publicly available at: https://github.com/BONS-AI-VCA-AMC/Endoscopy-CNNs-vs-Transformers.

Pre-training deep learning models with large data sets of natural images, such as ImageNet, has become the standard for endoscopic image analysis. This approach is generally superior to training from scratch, due to the scarcity of high-quality medical imagery and labels. However, it is still unknown whether the learned features on natural imagery provide an optimal starting point for the downstream medical endoscopic imaging tasks. Intuitively, pre-training with imagery closer to the target domain could lead to better-suited feature representations. This study evaluates whether leveraging in-domain pre-training in gastrointestinal endoscopic image analysis has potential benefits compared to pre-training on natural images. To this end, we present a dataset comprising of 5,014,174 gastrointestinal endoscopic images from eight different medical centers (GastroNet-5M), and exploit self-supervised learning with SimCLRv2, MoCov2 and DINO to learn relevant features for in-domain downstream tasks. The learned features are compared to features learned on natural images derived with multiple methods, and variable amounts of data and/or labels (e.g. Billion-scale semi-weakly supervised learning and supervised learning on ImageNet-21k). The effects of the evaluation is performed on five downstream data sets, particularly designed for a variety of gastrointestinal tasks, for example, GIANA for angiodyplsia detection and Kvasir-SEG for polyp segmentation. The findings indicate that self-supervised domain-specific pre-training, specifically using the DINO framework, results into better performing models compared to any supervised pre-training on natural images. On the ResNet50 and Vision-Transformer-small architectures, utilizing self-supervised in-domain pre-training with DINO leads to an average performance boost of 1.63% and 4.62%, respectively, on the downstream datasets. This improvement is measured against the best performance achieved through pre-training on natural images within any of the evaluated frameworks. Moreover, the in-domain pre-trained models also exhibit increased robustness against distortion perturbations (noise, contrast, blur, etc.), where the in-domain pre-trained ResNet50 and Vision-Transformer-small with DINO achieved on average 1.28% and 3.55% higher on the performance metrics, compared to the best performance found for pre-trained models on natural images. Overall, this study highlights the importance of in-domain pre-training for improving the generic nature, scalability and performance of deep learning for medical image analysis. The GastroNet-5M pre-trained weights are made publicly available in our repository: huggingface.co/tgwboers/GastroNet-5M_Pretrained_Weights.

Computer-aided detection and diagnosis systems (CADe/CADx) in endoscopy are commonly trained using high-quality imagery, which is not representative for the heterogeneous input typically encountered in clinical practice. In endoscopy, the image quality heavily relies on both the skills and experience of the endoscopist and the specifications of the system used for screening. Factors such as poor illumination, motion blur, and specific post-processing settings can significantly alter the quality and general appearance of these images. This so-called domain gap between the data used for developing the system and the data it encounters after deployment, and the impact it has on the performance of deep neural networks (DNNs) supportive endoscopic CAD systems remains largely unexplored. As many of such systems, for e.g. polyp detection, are already being rolled out in clinical practice, this poses severe patient risks in particularly community hospitals, where both the imaging equipment and experience are subject to considerable variation. Therefore, this study aims to evaluate the impact of this domain gap on the clinical performance of CADe/CADx for various endoscopic applications. For this, we leverage two publicly available data sets (KVASIR-SEG and GIANA) and two in-house data sets. We investigate the performance of commonly-used DNN architectures under synthetic, clinically calibrated image degradations and on a prospectively collected dataset including 342 endoscopic images of lower subjective quality. Additionally, we assess the influence of DNN architecture and complexity, data augmentation, and pretraining techniques for improved robustness. The results reveal a considerable decline in performance of 11.6% (±1.5) as compared to the reference, within the clinically calibrated boundaries of image degradations. Nevertheless, employing more advanced DNN architectures and self-supervised in-domain pre-training effectively mitigate this drop to 7.7% (±2.03). Additionally, these enhancements yield the highest performance on the manually collected test set including images with lower subjective quality. By comprehensively assessing the robustness of popular DNN architectures and training strategies across multiple datasets, this study provides valuable insights into their performance and limitations for endoscopic applications. The findings highlight the importance of including robustness evaluation when developing DNNs for endoscopy applications and propose strategies to mitigate performance loss.