Background & Summary

Gastric cancer (GC) has emerged as the fifth most commonly malignant tumor, presenting a significant global health challenge due to its high incidence and mortality1,2. Surgical resection remains the primary treatment for GC. Currently, minimally invasive surgery particularly laparoscopic and robot-assisted surgery have gradually become mainstream in surgical practice which offers several advantages over open surgery: smaller incisions, better surgical vision, less intraoperative blood loss, and more rapid postoperative recovery3,4,5. Notably, laparoscopic GC (LapGC) surgery has been proven by the Chinese Laparoscopic Gastrointestinal Surgery Study (CLASS)3,4,6, the Korean Laparoendoscopic Gastrointestinal Surgery Study (KLASS)7,8,9, and the Japan Clinical Oncology Group study (JCOG)10,11, a series of highquality clinical trials, to have better short-term outcomes and comparable long-term oncological outcomes relative to open surgery, which is recommended by several international treatment guidelines12,13.

According to guidelines, D2 lymph node dissection (LND) is the standard surgical procedure for patients with locally advanced GC, but it presents significant technical challenges, especially because of the complex and numerous perigastric vessels, which makes laparoscopic D2 LND more difficult13,14. The major difficulty of D2 LND lies in identifying the correct vascular sheath layers and dissecting lymphatic and fatty tissues to skeletonize vessels, which can easily result in vascular injuries and intraoperative bleeding15,16,17. A retrospective study has shown that the probability of vascular injury occurrence is 45% due to the intricate anatomy and vascular variations around the stomach18. Also, intraoperative bleeding caused by vascular injuries is one of the most common intraoperative complications19, which might influence the tumor patients’ prognosis severely20. Moreover, a recent study found that nearly 20% of intraoperative adverse events are caused by human visual perception errors in identification including anatomy misidentification21. Therefore, introducing innovative technology is crucial to assist surgeons in rapidly and accurately identifying and locating key vessels, which can effectively reduce the incidence of intraoperative vascular injury complications and enhance surgical safety.

Surgical data science (SDS) aims to optimize surgical outcomes by leveraging advanced data analytics and artificial intelligence (AI), including machine learning (ML), to improve decision-making22,23,24. There is a notably vibrant ___domain within SDS that incorporates the utilization of AI technologies, especially computer vision algorithms for the analysis of minimally invasive surgical videos and the development of AI applications in surgery25. These endeavours aim to furnish surgeons with context-aware intraoperative assistance, ultimately enabling more precise surgical treatment and improving clinical practice25. However, the development of surgical AI applications requires high-quality, publicly available image datasets that contain massive surgical information26, especially anatomical structures27,28. Intraoperative automatic recognition of anatomical structures, which is currently one of the most pivotal research focuses for surgeons, can be utilized for intelligent surgical navigation29 and assistance30, minimizing intraoperative complications risks31, identifying safety zones32, and enhancing surgical education33. At present, the publicly available surgical anatomy datasets with pixel-level segmentation annotations are still limited, which are summarized in Table 127,34,35,36,37,38,39,40,41, such as Dresden Surgical Anatomy Dataset27, LapGyn434, CholeSeg8k35, and Endoscapes41. These datasets have indeed promoted the development of SDS and surgical AI42. However, these existing publicly available surgical anatomy datasets primarily collect images of major organs within the abdominal cavity, covering relatively simple surgical procedures including gallbladder, gynecological, and colorectal surgeries. In particular, they lack detailed anatomical representations of the dynamic exposure, dissection and ligation process of key vessels in complex cancer surgery. Consequently, the generalizability and applicability of the above datasets are still limited for LapGC surgery, which involves more complex surgical scenes and vascular anatomy.

Table 1 Comparison of publicly available surgical anatomy datasets and the proposed Laparoscopic Gastric Cancer Key Vascular Anatomy Dataset.
Full size table

Thus, a key bottleneck currently hindering the development of AI applications in gastric cancer surgeries is the scarcity of high-quality datasets that incorporate detailed anatomical information, particularly concerning vascular structures. To meet this challenge and considering the specific surgical characteristics of D2 LND in GC surgery, this study proposed the LapGC Key Vascular Anatomy Dataset (LapGC-KVAD-30), also referred to as Southern Medical University - Nanfang Hospital - Gastric Cancer - 01 Dataset (SMU-NF-GC-01 Dataset), consisting of 5303 high-resolution frames from 30 complete videos across eight surgical scenes and fifteen types of key vessels, which are detailed in Table 2. The process of construction and annotation of the LapGC-KVAD-30 is illustrated in Fig. 1.

Table 2 The definition of key vessels during laparoscopic distal gastrectomy with D2 lymph node dissection.
Full size table
Fig. 1
figure 1

Overview of the Laparoscopic Gastric Cancer Key Vascular Anatomy Dataset. Based on the detailed definition of surgical scenes and key vessels in laparoscopic distal gastrectomy with D2 lymph node dissection, key frames were extracted by expert surgeons. Subsequently, medical annotators performed pixel-level annotations of the key frames under the supervision and guidance of expert surgeons. LapGC-KVAD-30: Laparoscopic Gastric Cancer Key Vascular Anatomy Dataset; SMU-NF-GC-01 Dataset: Southern Medical University-Nanfang Hospital-Gastric Cancer-01 Dataset.

Full size image

Compared to the publicly available surgical anatomy datasets listed in Table 1, the LapGC-KVAD-30 uniquely annotates the entire dynamic process of the exposure, dissection, or ligation of the key vessels crucial to D2 LND during GC surgery. Aimed at providing a foundational resource for the development of image-guided surgical AI navigation systems for GC, the LapGC-KVAD-30 enables surgeons and researchers to gain a deeper understanding of the anatomical relationships and vascular variations in LapGC surgery. This dataset can contribute to refining surgical planning, simulation training, and intraoperative navigation, ultimately enhancing the precision of LapGC surgeries and potentially improving clinical outcomes.

Methods

Dataset definition

In accordance with the guidelines for the treatment of GC, surgical anatomical characteristics and definition (Supplementary Note 1 and Table S1) of laparoscopic distal gastrectomy (LDG) with D2 LND, we systematically categorized the D2 LND procedure into eight critical scenes (S1–S8)12,13,43. Within these eight unique surgical scenes, the perigastric vessels requiring exposure, dissection, or ligation have been defined as “key vessels”, as detailed in Table 2 and Fig. 2. The temporal dynamic exposure, dissection, and ligation process of each type of key vessel is illustrated in Fig. 3 and as follows:

  • Left gastroepiploic LND (S1): In this scene, the surgeon dissects distally along the splenic vessels to expose the root of the left gastroepiploic artery (LGeA) and the left gastroepiploic vein (LGeV). The exposed vessels are then dissected, clipped, and cut, completing the No.4sb LND. The LGeA and the LGeV typically run in close proximity, with the LGeV generally positioned inferior to the LGeA. Therefore, the LGeA and LGeV (LGeA&V) are defined as accompanying vessels in this study.

  • Greater curvature LND (S2): In this scene, the surgeon cuts the vessels along the greater curvature and dissects the No.4d lymph nodes (LNs). The greater curvature vessels (GCVs) are defined as accompanying vessels, as the arteries and veins run in close proximity.

  • Infrapyloric LND (S3): In this scene, the surgeon exposes the right gastroepiploic vein (RGeV), the anterior superior pancreaticoduodenal vein (ASPDV), and the Accessory right colic vein (ARCV). The RGeV is dissected, clipped, and cut anterior to the APSDV. The surgeon then proceeds upwards to expose and carefully dissect the root of the right gastroepiploic artery (RGeA). Once the RGeA is fully exposed, it is clipped and cut at its root, where it originates from the gastroduodenal artery (GDA). At this point, the dissection of No. 6 LNs is completed.

  • Gastroduodenal artery exposure (S4): In this scene, the surgeon dissects the plane between the posterior wall of the duodenum and the pancreas to expose the GDA. Then, the surgeon dissects from the posterior to the superior wall of the duodenum, guided by the anterior space of the GDA.

  • Suprapyloric LND (S5): In this scene, the surgeon opens the anterior leaf of the hepatoduodenal ligament to expose the roots of the right gastric artery (RGA) and right gastric vein (RGV). The RGA and RGV are then dissected, clipped, and cut. The surgeon continues to dissect along the root of the RGA towards the liver, carefully exposing the anterior and medial walls of the proper hepatic artery (PHA), completing the No. 5 LND.

  • Suprapancreatic LND (S6): In this scene, the surgeon opens the gastropancreatic fold, locates the initial segment of the common hepatic artery (CHA) or splenic artery (SA), and dissects along the perivascular space towards the proximal segment of the celiac trunk. The left gastric artery (LGA) and left gastric vein (LGV) are exposed, dissected, clipped, and cut at their roots. Next, the lymphatic and fatty tissues on the lateral and posterior sides of the LGA are dissected, completing the dissection of the No. 7 and 9 LNs. The surgeon dissects along the superior, anterior, and posterior walls of the SA to complete the No.11p LND. Also, the surgeon dissects the CHA, exposing its anterior and superior walls to complete the No. 8a LND. Notably, during the process of No. 11p LND, the surgical quality control of the KLASS group44,45 standard requires that the splenic vein (SV) should be identified and exposed, or at least exposed the dorsal side of the pancreatic parenchyma. Li et al.46 suggested dissecting the fatty lymphatic tissue along the surface of the SA until near the posterior gastric artery (PGA) to complete No. 11p LND, without specifically requiring the exposure of the SV. In addition, the SV, located posterior to the SA, may exhibit vascular variations, making it occasionally invisible. Therefore, we did not define the SV as a type of key vessel due to the significant individual variability and uncertainty in its exposure during No. 11p LND.

  • No.12a LND (S7): In this scene, the surgeon performs blunt dissection along the potential anatomical space between the medial side of the PHA and the No. 12a LNs, exposing the portal vein (PV) posterior to the PHA. The lymphatic and fatty tissues from the anterior wall of the PV are then dissected, exposing the anterior and left walls of the PV, thus completing the No. 12a LND. Additionally, part of the inferior vena cava (IVC) may also be exposed during No.12a LND. The surgeon clips and cuts the root of the No. 12a LNs anterior to the IVC and PV, to prevent lymphatic leakage and avoid vascular injuries.

  • Lesser curvature LND (S8): In this scene, the surgeon cuts the vessels along the lesser curvature and dissects the No.1 and 3 LNs. The lesser curvature vessels (LCVs) are defined as accompanying vessels, as the arteries and veins run in close proximity.

  • Note: In this study, there are cases that lack certain types of key vessels due to following reasons: (1) vascular variations in patients, resulting in absence of specific key vessels; (2) complex intraoperative conditions (e.g., bleeding, smoke obstruction), which prevented the extraction of frames containing key vessels; (3) technical errors, which resulted in some key vessels not being properly exposed and thus invisible.

Fig. 2
figure 2

Anatomical illustration of surgical scenes and key vessels in laparoscopic distal gastrectomy with D2 lymph node dissection. In this figure, the labels of key vessels are highlighted in bold for emphasis. S1: Left gastroepiploic LND; S2: Greater curvature LND; S3: Infrapyloric LND; S4: Gastroduodenal artery exposure;S5: Suprapyloric LND; S6: Suprapancreatic LND; S7: No.12a LND; S8: Lesser curvature LND. LND: lymph node dissection; LGeA: left gastroepiploic artery; LGeV: left gastroepiploic vein; SA: splenic artery; SV: splenic vein; GCVs: greater curvature vessels; RGeA: right gastroepiploic artery; RGeV: right gastroepiploic vein; ASPDV: anterior superior pancreaticoduodenal vein; ARCV: accessory right colic vein; GCT: gastrocolic trunk; GDA: gastroduodenal artery; CHA: common hepatic artery; PHA: proper hepatic artery; RGA: right gastric artery; RGV: right gastric vein; LHA: left hepatic artery; RHA: right hepatic artery; LGA: left gastric artery; LGV: left gastric vein; PV: portal vein; IVC: inferior vena cava; LCVs: lesser curvature vessels.

Full size image
Fig. 3
figure 3

The exposure, dissection, and ligation process of key vessels in each scene in laparoscopic distal gastrectomy with D2 lymph node dissection. Each row represents a specific type of key vessel, along with the corresponding surgical steps for exposure, dissection, or ligation. LND: lymph node dissection; LGeA&V: left gastroepiploic artery&vein; GCVs: greater curvature vessels; RGeA: right gastroepiploic artery; RGeV: right gastroepiploic vein; GDA: gastroduodenal artery; PHA: proper hepatic artery; RGA: right gastric artery; RGV: right gastric vein; LGA: left gastric artery; LGV: left gastric vein; CHA: common hepatic artery; SA: splenic artery; PV: portal vein; IVC: inferior vena cava; LCVs: lesser curvature vessels.

Full size image

Data collection

Considering the time required and workload of frame extraction and annotation, we selected 30 complete LDG videos with D2 LND collected at Nanfang Hospital, Southern Medical University. To ensure the quality of dataset, we excluded low-quality videos, such as those with excessive bleeding, heavy smoke, or blurry clips. Also, this sample size is comparable to publicly available high-quality surgical anatomy datasets (Table 1)27,37,38,40,41. All video data were recorded with laparoscopic cameras from KARL STORZ and saved at a resolution of 1920 × 1080 pixels in MPEG-4 format. All identifiable patient information (e.g., names, ages, or other sensitive data) was anonymized. Also, each video was assigned a simple numeric identifier (ID) (e.g., 1, 2, 3) instead of using any patient-related IDs. This study aims to collect laparoscopic surgery videos and extract surgical images to build a dataset. Therefore, we applied for retrospective and observational ethical approval from the Medical Ethics Committee of Nanfang Hospital, Southern Medical University (Approval Number: NFEC-2023-467). Given the retrospective and observational nature of the study, and since all patient data were anonymized during the research process, the committee waived the requirement for individual patient consent.

Frame extraction

In this study, video frames were manually selected at a constant rate of 25 frames per second (fps) using the FFmpeg framework (www.ffmpeg.org) by three expert surgeons (H.C., Y.H., and J.Y.) who are core members of the CLASS group. The frame selection was based on the complexity of the vascular anatomy, surgical duration, and difficulty of surgical manipulation, resulting in a varying number of frames across different scenes and key vessels. Each frame was saved at a resolution of 1920 × 1080 pixels in PNG format. Frames were selected based on clarity and their temporal sequence to record the dynamic exposure, dissection, and ligation process of key vessels across eight surgical scenes. The following exclusion criteria were applied to frame selection: (1) frames containing any irrelevant operations or extra-abdominal operations; (2) unclear frames containing artifacts, glare, or smoke; (3) blurred or out-of-focus frames. To maintain the authenticity of frames containing vascular structures, no image pre-processing steps (e.g., adjusting image intensity, contrast, or window size) were performed.

Segmentation annotation

The annotation group consisted of three expert surgeons (H.C., Y.H., and J.Y.) with experience in over 300 LapGC surgeries and five medical annotators including a novice surgeon (L.G.) and four medical students (H.W., C.C., B.S., and Y.Q.). The novice surgeon had completed a rotation in the surgical department and participated in over 20 LapGC surgeries. The four medical students are those who have completed undergraduate medical courses including anatomy and surgery, and have watched at least 30 LapGC surgery videos. Strict and standardized training was an essential part and cornerstone of the annotation process47. At first, all members of the annotation group systematically learned how to use the annotation tool LabelMe48. Then, the five medical annotators completed training including learning the process of LDG with D2 LND and identifying each type of key vessel. Finally, we developed a segmentation annotation protocol as a guideline that details the annotation criteria for each type of key vessel across eight different scenes (Supplementary Note 2, Table S2, and Figure S1–S14). Frames from the same video were assigned to a single annotator for pixel-level segmentation annotation. Medical annotators used the polygonal annotation tool in LabelMe to outline the boundaries of the key vessels, adhering strictly to the annotation protocol. Then, the three expert surgeons conducted a thorough review of annotations together to correct any errors, such as misidentification of blood-bearing adipose tissue as key vessels. Annotations not meeting the standards outlined in our annotation protocol were sent back to the medical annotators for re-annotation. Sample annotation frames of each type of key vessel across eight scenes are presented in Fig. 4.

Fig. 4
figure 4

Sample annotation frames for each type of key vessel across eight scenes. The figure displays raw frames (left column) and their corresponding pixel-wise annotated frames (right column) for key vessels in D2 lymph node dissection during laparoscopic distal gastrectomy. LGeA&V: left gastroepiploic artery&vein; GCVs: greater curvature vessels; RGeA: right gastroepiploic artery; RGeV: right gastroepiploic vein; GDA: gastroduodenal artery; PHA: proper hepatic artery; RGA: right gastric artery; RGV: right gastric vein; LGA: left gastric artery; LGV: left gastric vein; CHA: common hepatic artery; SA: splenic artery; PV: portal vein; IVC: inferior vena cava; LCVs: lesser curvature vessels.

Full size image

Data Records

The LapGC-KVAD-30 is stored at the Science Data Bank (https://doi.org/10.57760/sciencedb.20852)49. This dataset consists of three independent subsets: the training set (20 cases), validation set (5 cases), and test set (5 cases). Each folder corresponds to a single case. All folders within these subsets share the same folder structure, which is illustrated in Fig. 5 and described below:

  • Folders: Each folder is named according to the unique ID number of its corresponding surgical video, ranging from “1” to “30”. Each folder contains 8 subfolders named “S1”, “S2”,…, “S8” and 1 CSV file.

    • Subfolders: Each subfolder corresponds to a specific surgical scene and contains two types of PNG files, which are stored in two separate folders: “frames” and “annotations”. In “frames” folder, the filenames follow a structured format that includes the surgical video ID, scene ID, and the frame sequence number. For example, a filename like “10_S3_25.png” indicates that this file comes from a surgical video whose ID is number 10, specifically from the 25th frame selected from Scene 3 (S3) within this video. The “annotations” folder contains the corresponding segmentation mask files, which are saved in PNG format. Compared to the frame files, each mask file has a “_mask” suffix. For example, the corresponding mask file for “10_S3_25.png” is named “10_S3_25_mask.png”.

    • CSV file: The CSV file provides a detailed record that includes the filenames of the frames, their corresponding mask filenames, the video ID, the scene ID, and the names of the key vessels present.

  • Color file: The “color file” is a JSON-format file that stores the names of fifteen types of key vessels, the Red-Green-Blue (RGB) values of their corresponding colors as displayed in the segmentation masks, and the names of their associated scenes.

Fig. 5
figure 5

Dataset folder structure. LapGC-KVAD-30 has three subsets (training, validation, and test) and a file named “color.json”. The training, validation, and test sets contain 20, 5, and 5 folders respectively. Each folder corresponds to a single case and contains eight subfolders, which include frames, segmentation annotations, and a CSV file that stores the names of all files along with their corresponding information. The “color.json” file stores the RGB values of the colors for fifteen types of key vessels. For better understanding, we visualized the “color.json” file as a table and displayed the color corresponding to each RGB value. LapGC-KVAD-30: Laparoscopic Gastric Cancer Key Vascular Anatomy Dataset; SMU-NF-GC-01 Dataset: Southern Medical University-Nanfang Hospital-Gastric Cancer-01 Dataset; RGB: Red-Green-Blue. ID: identifier.

Full size image

Technical Validation

Dataset characteristics

The distribution of frames across the eight critical surgical scenes in the LapGC-KVAD-30 is visualized in Fig. 6a, with S1 to S8 accounting for 12.60%, 7.98%, 23.49%, 6.56%, 11.79%, 23.61%, 6.75% and 7.22% of the total frames, respectively. Figure 6b shows the frame count at which key vessels appear across these eight critical scenes, providing a thorough visualization that reveals the distribution patterns of different types of key vessels in each scene. Figure 6c illustrates the distribution of these eight scenes across all 30 videos, including the total frame count of each video and the frame count of each scene within the videos. Overall, the distribution of frame count in the LapGC-KVAD-30 varies across the eight critical scenes, with notable differences in both the frame count and distribution patterns of key vessels in each scene. This highlights the complexity of vascular anatomy during the LapGC surgery and the critical importance of constructing a key vascular anatomy dataset. In addition, Fig. 6d showed that the frame proportion of different scenes in the training set, validation set, and test set is similar to that in the entire dataset, indicating that the data separation is appropriate, which is beneficial for future benchmarking.

Fig. 6
figure 6

Visualization of frame count distribution for Laparoscopic Gastric Cancer Key Vascular Anatomy Dataset. (a) The distribution of the 5303 key frames across the eight scenes (S1-S8) in LapGC-KVAD-30. (b) The count of frames for each key vessel across the eight scenes. Vessels are categorized by artery (red), vein (blue), and accompanying vessels (green). (c) The frame count for each of the eight scenes per video. (d) The frame proportion of eight scenes in LapGC-KVAD-30 and its three subsets. LGeA&V: left gastroepiploic artery&vein; GCVs: greater curvature vessels; RGeA: right gastroepiploic artery; RGeV: right gastroepiploic vein; GDA: gastroduodenal artery; PHA: proper hepatic artery; RGA: right gastric artery; RGV: right gastric vein; LGA: left gastric artery; LGV: left gastric vein; CHA: common hepatic artery; SA: splenic artery; PV: portal vein; IVC: inferior vena cava; LCVs: lesser curvature vessels; ID: identifier.

Full size image

Intra- and inter-annotator consistency

In this study, frames were annotated by five medical annotators and refined by three expert surgeons. Thus, three kinds of consistencies should be evaluated: intra-annotator consistency (same annotator at different times), inter-annotator consistency (medical annotator vs. medical annotator), and inter-annotator consistency (medical annotator vs. expert surgeon). We used these metrics to calculate the three kinds of consistencies27:

  • Dice coefficient (Fig. 7a): The Dice coefficient is a metric that measures the degree of overlap between distinct annotations, ranging from 0 to 1, where 0 indicates no overlap at all, and 1 signifies complete overlap.

    Fig. 7
    figure 7

    Consistency evaluation metrics. (a) The formula for the Dice coefficient. (b) The formula for the normalized 95% Hausdorff Distance (95%HDnorm). “X” denotes the segmentation annotation of annotator-1, while “Y” refers to the annotation of annotator-2. The symbol “ ∩ ” represents the intersection operation between X and Y. “dXY” refers to the minimum Euclidean distance from a point \(x\in X\) to the set Y. “dYX” refers to the minimum Euclidean distance from a point y \(\,\in Y\) to the set X.

    Full size image

  • The 95th Hausdorff Distance (Fig. 7b): The 95th Hausdorff Distance (95% HD) is a robust distance metric used to evaluate the dissimilarity between a reference annotation and another annotation. It computes the maximum of the directed distances from each point in one set to the nearest point in the other set, taking into account the 95th percentile to reduce the influence of outliers. In this study, we normalize this metric relative to the length of the frame diagonal, masking values within the range of 0 to 1. A value of 0 in the normalized 95%HD (95% HDnorm) denotes perfect correspondence between the annotations (zero separation), while a value of 1 represents the extreme case where the greatest distance between any matched points equals the image diagonal.

A subset containing 100 frames was randomly selected from LapGC-KVAD-30 for consistency evaluation, covering all scenes and key vessel types. The analysis aimed to evaluate the three kinds of consistencies, which were detailed and displayed in Table 3, Table S3 and as follows:

  • For intra-annotator consistency analysis of the same annotator at different times, five medical annotators (L.G., H.W., C.C., Y.Q., and B.S.) from the annotation group were asked to annotate the 100 images twice within one month, with at least one 1-week interval between each annotation. The mean Dice coefficients for different types of vessels varied from 0.93 to 0.97, suggesting high consistency. Additionally, 95% HDnorm values under 0.01 corroborated the strong intra-annotator reliability.

  • For inter-annotator consistency analysis of the medical annotators, the annotation results of five medical annotators (L.G., H.W., C.C., Y.Q., and B.S.) were compared. The mean Dice coefficients exhibited a slight decrease, with values ranging from 0.85 to 0.94, suggesting a satisfactory level of consistency among annotators of the same expertise level. Correspondingly, the 95% HDnorm values between 0.00 and 0.04 also supported this satisfactory consistency.

  • For inter-annotator consistency analysis of medical annotators and the expert surgeon, the expert surgeon (H.C.) annotated the 100 frames for comparison with five annotators (L.G., H.W., C.C., Y.Q., and B.S.). The mean Dice coefficients ranged from 0.84 to 0.93, showing a close alignment between annotators and the expert surgeon, which indicates a consistent annotation approach across different expertise levels after strict training. The 95% HDnorm values, which fell within the range of 0.01 to 0.04, echoed this finding.

Table 3 Dice coefficient for intra-annotator consistency and inter-annotator consistency between medical annotators and the expert surgeon.
Full size table

In conclusion, the consistency analysis of the subset confirmed that annotations were consistently reliable regardless of the annotator, the time of annotation, and the annotators’ level. This high level of consistency lays a solid foundation for accurate and reproducible segmentation of key vessels. Similarly, the public Dresden surgical anatomy datasets27 reported Dice coefficients ranging from 0.62 to 0.99 when evaluating the annotation consistency between different annotators, which aligns with the results of our study. This underscores the importance of strict annotation training, expert review, and refinement of annotations in the construction of high-quality surgical anatomy datasets.

Advantages and limitations of the dataset

Currently, publicly available surgical anatomy datasets mainly focus on gallbladder, gynecological, and colorectal surgeries27,34,41, with no data available for GC surgery. These datasets typically focus on organs and nearby structures (e.g. liver, pancreas, cystic artery), lacking images showing the manipulation of key vessels during surgery. However, GC surgery is challenging due to the complex anatomy and the numerous perigastric vessels. During D2 LND, it is crucial to dissect, expose, or ligate key vessels in different scenes. This dataset contains images of dynamic manipulation processes of perigastric vessels in LDG with D2 LND, which will help advance AI research in GC surgery and improve surgical safety.

However, this dataset faces several limitations. Firstly, to construct a high-quality dynamic vascular anatomy dataset, we selected surgical videos with minimal bleeding, limited smoke, and clear frames. As a result, the current dataset may exhibit a selection bias based on the clarity of the surgical frames. Secondly, due to the inherent differences in complexity of vascular anatomy, surgical duration of scenes, and the difficulty of surgical manipulations, the number of frames containing key vessels in each scene may be different. Thus, there are class imbalances of surgical scenes and key vessels causing by these differences. However, the class imbalances accurately reflect the the complexity of perigastric vascular anatomy and technical characteristics of D2 LND for GC. Thirdly, the frames in this dataset were recorded using the same type of laparoscopic device at a single center, with a white-light modality. Therefore, the AI models developed based on this dataset may have limited generalization performance when applied to different centers, laparoscopic devices, and imaging modalities. In the future, potential solutions to above limitations include collecting surgical videos with varying levels of clarity and gathering multi-center, multi-device, and multi modality imaging data to increase the diversity of the dataset and reduce selection bias. Especially, data augmentation50 and customized weighted loss functions51 can be applied to address class imbalance and enhance the performance of AI models.