This is a multicenter study with two phases. During the first observational phase conducted from February, 2023 to February, 2024, models were trained using retrospectively collected videos. A total of 581 videos of 541 laparoscopic and 40 robotic pancreatic surgeries were collected from five clinical centers (Peking Union Medical College Hospital, PUMCH; Peking University Third Hospital, PUH3; the Xijing Hospital of the Fourth Military Medical University, XJH; Fudan University Shanghai Cancer Center, FUSCC; and Tongji Hospital, WHTH). The inclusion criteria were as follows: (1) laparoscopic or robotic pancreatic surgery; (2) gauze appearing in the video data for ≥ 3 min. Exclusion criteria included: (1) any missing fragment in videos; (2) conversion to open surgery or termination of surgery for any reason. During the second interventional phase conducted from March, 2024 to May, 2024, after the model was developed, we initiated an exploratory clinical application in the surgical rooms of PUMCH and XJH. Five patients were included in each center. Three video clips were collected for each patient, generating 4,832 images for PUMCH and 4,672 images for XJH. Data of XJH generated during this process was used as an independent real-patient test set [Supplementary Table 1].

During an early pilot exploration, we evaluated sampling one frame every 24 frames to assess frame redundancy and annotation workload. This exploratory sampling strategy was discontinued at an early stage and was not used for construction of the final dataset or for model training and evaluation. For the final dataset, one frame was sampled every four frames from the selected video segments. The same sampling strategy was applied consistently across the datasets used for model development and evaluation.

Each selected frame was independently annotated by two annotators according to the same annotation protocol. The smallest rectangle that can encompass the target gauze was labeled. Four annotators with at least one year of surgical training participated in the annotation process. The annotations were subsequently reviewed by a senior surgeon with extensive experience in pancreatic surgery. Discrepant or ambiguous annotations were re-examined and resolved under the supervision of the senior surgeon before inclusion in the final dataset. The annotated images were divided into training and test sets separately for each clinical center on a patient level. To prevent information leakage, all frames extracted from the same surgical video were assigned exclusively to either the training or test set. When multiple videos were available from the same patient, all videos from that patient were assigned to the same partition. The allocation was performed before frame extraction. The final numbers of training and test images were as follows: PUMCH, 160,390 and 15,071; XJH, 43,734 and 4,545; WHTH, 20,474 and 5,255; FUSCC, 15,648 and 5,575; and PUH3, 20,162 and 5,273, respectively. The center-specific allocation of images is summarized in Table 1. No test images were used for model training.

We developed a lightweight modified detector based on the YOLOv3 framework [Figure 1]. The backbone extracted features at three spatial scales, corresponding to feature-map strides of 32, 16, and 8. These feature maps were subsequently processed by a YOLOv3 neck and detection head to support multiscale gauze detection. Three modifications were introduced during model development. First, deconvolution was used to enlarge feature maps and facilitate feature fusion across different scales. Second, max-pooling was used to reduce computational cost. Third, channel pruning was performed after completion of model training to decrease the number of parameters and computational complexity. We used an AutoML-based pruning scheme in which a reinforcement-learning algorithm automatically searched for the pruning ratio of each network layer. The performance-difference threshold between the pruned model and the original trained model was set at 1%. These modifications were selected a priori with the practical design objective of reducing computational burden and facilitating real-time deployment while retaining the multiscale detection capability of the YOLOv3 framework.

The backbone provided feature outputs with channel dimensions of 1024, 512, and 256 at strides of 32, 16, and 8, respectively. These features were processed by the neck, producing corresponding output channel dimensions of 512, 256, and 128. The anchor generator used three groups of anchor sizes corresponding to the three detection scales: (116, 90), (156, 198), and (373, 326) for the stride-32 feature map; (30, 61), (62, 45), and (59, 119) for the stride-16 feature map; and (10, 13), (16, 30), and (33, 23) for the stride-8 feature map.

The model was trained using stochastic gradient descent with a learning rate of 0.002, momentum of 0.9, and weight decay of 0.0005. The batch size was 16, and the training schedule consisted of a maximum of 100 epochs. No separate validation set was used in this study. The center-specific test sets were separated at the patient and video levels before frame extraction and were not used for model training or checkpoint selection. The final checkpoint was selected as the checkpoint obtained at the end of the prespecified 100-epoch training schedule.

During training, positive and negative samples were assigned using a GridAssigner with positive and negative IoU thresholds of 0.5. During testing, predictions with a confidence score below 0.45 were discarded. Non-maximum suppression (NMS) was then applied using an IoU threshold of 0.3. A maximum of 100 detections was retained per image.

The Intersection over Union (IoU) is defined as:

$$ \mathrm{IoU}=\frac{\operatorname{area}(\mathrm{P}) \cap \operatorname{area}(\mathrm{GT})}{\operatorname{area}(\mathrm{P}) \cup \operatorname{area}(\mathrm{GT})} $$

Here, P denotes the predicted bounding box and GT denotes the ground-truth bounding box.

Predictions with a confidence score below 0.45 were discarded. NMS was then applied using an IoU threshold of 0.3. After NMS, each ground-truth box was matched to at most one prediction. A prediction with IoU > 0.3 with an unmatched ground-truth box was counted as a true positive; unmatched predictions were counted as false positives, including residual duplicate predictions; and unmatched ground-truth boxes were counted as false negatives. We selected this threshold because the primary objective of the model was to identify and continuously track the presence and approximate location of surgical gauze during laparoscopic procedures, rather than to achieve highly accurate contour localization. This criterion was predefined before model development and was consistently applied throughout all experiments.

Sensitivity rate (SR) denotes the proportion of ground-truth targets that are correctly detected, whereas precision rate (PR) denotes the proportion of correct predictions among all predictions, which are defined as follows:

$$ \text { Sensitivity rate }(\mathrm{SR})=\frac{\text { True Positive }}{\text { True Positive }+ \text { False Negative }} $$

$$ \text { Precision rate (PR) }=\frac{\text { True Positive }}{\text { True Positive + False Positive }} $$

False negative rate (FNR) refers to the proportion of positive samples that are not predicted correctly, and false discovery rate (FDR) refers to the proportion of false-positive predictions among all positive predictions. FNR and FDR were calculated by:

$$ \text { False negative rate }(\mathrm{FNR})=1-\mathrm{SR} $$

$$ \text { False discovery rate }(\mathrm{FDR})=1-\mathrm{PR} $$

The performance of all models generated was tested in test sets from all centers separately and combined.

The in vitro test environment of the model is PyTorch: 1.11-cuda11.3-cudnn8-ubuntu20.04. The environment configurations are: Torch v1.11, MMEngine 0.6.0, MMCV 2.0.0rc4, MMDetection 3.0.0rc6, MMYOLO 0.5.0.

The baseline model was constructed first on data of PUMCH. SR, PR, FNR, FDR, and running speed were evaluated.

To investigate the appropriate sample size for stabilizing the performance of the model in the same scene, we first collected data from XJH for exploration. The dataset from XJH was generated following the method described in the Section “Video annotation and dataset division”. Starting from the model trained on the base training set, we added 1,000, 5,000, 10,000, 15,000, 20,000, 25,000, 30,000, 35,000, and 40,000 images from XJH to the training set [Table 1]. Changes in sensitivity and precision were then evaluated.

Due to differences in surgical style and image characteristics among hospitals, model performance may vary across clinical centers. We therefore used the PUMCH-trained baseline model as the starting point and sequentially incorporated training data from the other four centers to assess cross-center performance variation and the effect of additional center-specific data on sensitivity and precision. For each hospital, approximately 5,000 images were selected as the test set, and the remaining images were selected as the training set. The exact numbers of training and test images used for each center are provided in Table 1 and its footnote. This amount was based on the experimental results of labeling cost and data volume. To optimize computational efficiency, the strategy of adding training set data from XJH, WHTH, FUSCC and PUH3 was adopted. This allowed us to monitor changes in sensitivity and precision between the old model and the new model in the test set after each training iteration.

The study was performed in accordance with the Declaration of Helsinki. The study was approved by the ethics committees of the participating hospitals: I-23PJ253 (the Institutional Review Board of Peking Union Medical College Hospital, which reviewed the conduction of the study in all centers) & I-23PJ343 (the Institutional Review Board of Peking Union Medical College Hospital, which reviewed the conduction of the study at PUMCH); IRB00006761-M2023135 (Peking University Third Hospital Medical Science Research Ethics Committee); KY20232167-F-1 (the Medical Ethics Committee of the First Affiliated Hospital of the Air Force Medical University); 2303272-6 (Shanghai Cancer Center Institutional Review Board); TJ-IRB20230420 (Institutional Review Board, Tongji Hospital, Tongji Medical College, Huazhong University of Science and Technology). At PUH3, FUSCC and WHTH, informed consent was waived as surgical videos were collected only retrospectively with no identifiable information included. At PUMCH and XJH, written informed consent was obtained from each participant before the collection of intraoperative videos, intraoperative photographs, and medical-record data. For the retrospective observational phase, an institutional consent form that authorized the collection and use of these materials for clinical teaching and medical research, with appropriate protection of personal privacy, was used. The consent form was an institutional broad informed consent document and was not specific to the present manuscript. For the interventional phase, an additional informed consent was obtained from all participants to inform them of the details of the study.

581 videos with a total of 296,127 frames of images were labeled [Table 1]. Among the different clinical centers, PUMCH and XJH were included in the data volume experiment, whereas all centers were included in the generalization experiment.

The model trained by the baseline dataset had satisfactory results of detecting gauze under different situations including image blur, irregular shape, and blood immersion [Figure 2]. The baseline model, trained using the PUMCH training set, achieved an SR of 87.46% and a PR of 88.28% on the PUMCH test set. When evaluated on the combined test set from all five centers, the corresponding SR and PR were 84.70% and 79.58%, respectively [Table 2]. The inference time was 29.04 ms per frame.

In order to explore the robustness of our model across different datasets, we first directly applied our baseline model to the XJH test set. The SR and PR of the baseline model in the XJH test set were 83.30% and 68.79%, respectively, both of which were lower than the corresponding values on the PUMCH test set [Table 2]. We attempted to enhance the performance of the model on the XJH test set by sequentially adding 0 to 40,000 images from the XJH training set into our training sequence, and SR of the model on the XJH test set increased by 7.31 percentage points with the addition of data, whereas PR varied within a relatively narrow range [Table 3]. In addition, adding more XJH images did not adversely affect model performance on the PUMCH test set [Table 3]. However, the increase in sensitivity was not consistently accompanied by an increase in precision. In particular, on the XJH test set, sensitivity increased to 90.61% after additional XJH training data were incorporated, whereas precision remained approximately 68%-70%. Still, solely enlarging the training set could not improve model performance indefinitely. There was a slowdown in the growth of SR after 25,000 images from the XJH training set were added [Figure 3].

After demonstrating the beneficial effect of enlarging the training set, we then sought to examine the generalization potential of our model by sequentially adding more clinical centers to the training set. We first evaluated the PUMCH-trained baseline model on independent test sets from the other four clinical centers, none of which contributed training data to Model 1. The sensitivity and precision values varied across centers, indicating differences in model performance under different center-specific data distributions. We then progressively incorporated training data from the additional centers and evaluated the resulting models on the predefined test sets. Addition of center-specific training data generally improved sensitivity on the corresponding test sets, whereas changes in precision were variable [Table 2]. The SR/PR of baseline model on the merged test sets was 84.70%/79.58%. On independent test sets from four hospitals, the sensitivity/precision was 83.30%/68.79%, 85.50%/77.26%, 88.28%/76.34%, and 75.17%/78.83%, respectively. After incorporating the training data from each hospital, the sensitivity/precision on their respective test sets improved by 4.90/-0.71, 5.59/3.39, 5.03/3.15, and 16.93/-2.62 percentage points compared to the baseline model. Sensitivity and precision also improved on the overall test set [Table 2].

After completing the above tests, we have initiated an exploratory clinical attempt and used this model in actual patients [Figure 4, Supplementary Table 1 and Supplementary Video 1]. In the preliminary clinical use, the model provided real-time gauze localization without subjectively noticeable delay. Surgeons provided generally positive qualitative feedback.

We extracted a part of the images from the total test set for the analyses of missed detection and false detection [Table 4]. We found that, in these images, the reasons for missed detection included: (1) Target objects were blocked by tissues or instruments and only a small part of gauze was exposed [Figure 5A]; (2) Target objects received poor illuminance due to viewing angle problems [Figure 5B]; (3) Target objects were severely soaked in blood and lost part of their texture features [Figure 5C]. The reasons for misdetection include target segmentation [Figure 6A] and misdetection caused by tissues and instruments with similar image features [Figure 6B].

The modified YOLOv3-based model enabled real-time gauze detection in the evaluated multicenter surgical datasets. Incorporating data from multiple centers was associated with improved performance on the corresponding test sets, supporting the feasibility of multicenter training for exploratory cross-center evaluation. Further prospective studies involving completely unseen centers and predefined clinical endpoints are needed to determine its clinical benefit and effect on retained foreign-body prevention.