Deep-Learning Tool Improves Anatomical Recognition in Pelvic Surgery

An artificial intelligence model (AI) has been designed to identify key anatomical structures during pelvic lymph node dissection (PLND). Trained on surgical videos, the model achieved strong accuracy for the obturator nerve and the two major iliac vessels, though recognition of the ureter proved more difficult. 

Despite these difficulties, AI assistance significantly improved surgeons' recognition of all four structures across specialties and experience levels, suggesting potential to enhance surgical safety. These findings were published in Npj Digital Medicine.

Futuristic skull scan being analyzed by a female scientist
Study: Enhancing anatomical recognition by surgeons during pelvic lymph node dissection using artificial intelligence. Image Credit: PeopleImage/Shutterstock.com

The Need for AI-Assisted Navigation in Pelvic Surgery

PLND is a critical procedure in the surgical management of colorectal, gynecological, and urological cancers. However, it remains technically challenging due to the complex pelvic anatomy and the proximity of vital structures such as ureters, nerves, and major vessels, which carry a significant risk of injury.

While deep-learning-based semantic segmentation has shown promise for intraoperative guidance in minimally invasive surgery, this study is among the first to apply the approach specifically to PLND using a multidisciplinary dataset.

The researchers reasoned that training on data spanning colorectal, gynecological, and urological procedures would improve the model's recognition performance and generalizability compared to a single-specialty dataset.

Saving this for later? Download a free PDF now!

Model Development and Evaluation Approach

This multicenter, prospective observational study evaluated PelviX3Net, a deep-learning-based semantic segmentation model developed to automatically recognize key anatomical structures during minimally invasive pelvic lymph node dissection, including the ureter, obturator nerve, external iliac artery, and external iliac vein.

The model employed a UNet++ architecture with an EfficientNetV2-L backbone pretrained on ImageNet-21K. Anonymized videos of laparoscopic and robot-assisted procedures were converted to static images for training, with no clinical metadata or patient demographics available due to complete anonymization.

Twelve annotators, four colorectal surgeons, five gynecologists, and three urologists generated ground-truth labels using the computer vision annotation tool. All annotations were double-checked, and unannotated images without target structures were included as negative training data to reduce false positives.

For validation, three-fold cross-validation using the Dice similarity coefficient (DSC) was performed. The test dataset comprised 30 cases from multiple hospitals across Japan, which were 10 rectal resections, 10 total hysterectomies, and 10 radical prostatectomies, with patient sex and camera system manufacturers adjusted to reflect real-world distribution.

From each video, eight half-second snippets with target structures and eight without were randomly extracted, yielding 480 snippets. Detection performance was evaluated using sensitivity and specificity, with true positive defined as achieving DSC above predetermined structure-specific thresholds on five or more consecutive frames.

Thirty-six surgeons across colorectal surgery, gynecology, and urology, ranging from non-board-certified trainees to highly specialized certified surgeons with up to 25 years of experience, participated in the evaluation. Surgeons viewed 320 snippets per condition in a side-by-side monitor setup, plotting points on the last frame when a target structure was identified.

The primary endpoint was sensitivity of structure recognition with versus without AI assistance, and specificity was the secondary endpoint, with subgroup analyses comparing board-certified specialists and trainees. Statistical comparisons used Student's t-test, with significance set at p<.05.

Enhancing Surgical Vision Through AI Assistance

The training dataset for PelviX3Net comprised 23,259 annotated and 653 unannotated images extracted from 293 minimally invasive PLND videos, stratified by surgical specialty.

Colorectal surgeons contributed 128 videos yielding 8546 annotated and 197 unannotated images; gynecologists contributed 121 videos yielding 9936 annotated and 323 unannotated images; and urologists contributed 44 videos yielding 4777 annotated and 133 unannotated images.

Three-fold cross-validation performed on this dataset demonstrated DSC of 0.6483 for the ureter, 0.8654 for the obturator nerve, 0.8619 for the external iliac artery, and 0.8736 for the external iliac vein.

On the independent test dataset, recognition performance revealed that the ureter showed relatively low sensitivity at 62.9% while maintaining high specificity at 96.3%, whereas the obturator nerve, external iliac artery, and external iliac vein achieved consistently high sensitivity and specificity, ranging from 86.3% to 96.3%.

Regarding surgeon performance, AI assistance resulted in significantly improved sensitivity for all four target structures across all participating surgeons. Subgroup analysis by expertise level demonstrated that both board-certified specialists and trainees derived significant benefit from AI assistance for most target structures. 

The only exception to this was trainee recognition of the external iliac vein, which showed no significant improvement.

For specificity, AI assistance provided significant additional benefits across all target structures for participating surgeons, board-certified specialists, and trainees alike. Notably, the improvement in specificity with AI assistance was particularly significant among trainees, indicating that the AI effectively aided in minimizing misidentification among less experienced surgeons.

These findings collectively demonstrate that the developed model significantly enhanced surgeons' ability to correctly identify target anatomical structures while simultaneously reducing misrecognition, with inexperienced trainees deriving especially substantial benefit.

Toward Safer Pelvic Surgery with AI

This study developed an AI model that reliably identified three of its four target structures, though ureter recognition lagged behind and remains a limitation the authors flag as needing further work before routine clinical use.

AI assistance significantly improved both sensitivity and specificity of anatomical recognition among surgeons across specialties and experience levels, with trainees deriving particularly notable benefit in reducing misidentifications.

The authors describe the improvement as moderate overall, cautioning that the study was observational, tested surgeons on short, context-free video clips rather than live surgery, and did not evaluate the model in an actual operating room.

Future studies using continuous intraoperative workflows will be needed to determine whether these gains translate into real clinical impact.

Journal Reference

Kitaguchi, D. et al. (2026). Enhancing anatomical recognition by surgeons during pelvic lymph node dissection using artificial intelligence. Npj Digital Medicinehttps://www.nature.com/articles/s41746-026-02936-4.

Disclaimer: The views expressed here are those of the author expressed in their private capacity and do not necessarily represent the views of AZoM.com Limited T/A AZoNetwork the owner and operator of this website. This disclaimer forms part of the Terms and conditions of use of this website.

Citations

Please use one of the following formats to cite this article in your essay, paper or report:

  • APA

    Nandi, Soham. (2026, July 07). Deep-Learning Tool Improves Anatomical Recognition in Pelvic Surgery. AZoRobotics. Retrieved on July 07, 2026 from https://www.azorobotics.com/News.aspx?newsID=16435.

  • MLA

    Nandi, Soham. "Deep-Learning Tool Improves Anatomical Recognition in Pelvic Surgery". AZoRobotics. 07 July 2026. <https://www.azorobotics.com/News.aspx?newsID=16435>.

  • Chicago

    Nandi, Soham. "Deep-Learning Tool Improves Anatomical Recognition in Pelvic Surgery". AZoRobotics. https://www.azorobotics.com/News.aspx?newsID=16435. (accessed July 07, 2026).

  • Harvard

    Nandi, Soham. 2026. Deep-Learning Tool Improves Anatomical Recognition in Pelvic Surgery. AZoRobotics, viewed 07 July 2026, https://www.azorobotics.com/News.aspx?newsID=16435.

Tell Us What You Think

Do you have a review, update or anything you would like to add to this news story?

Leave your feedback
Your comment type
Submit

Sign in to keep reading

We're committed to providing free access to quality science. By registering and providing insight into your preferences you're joining a community of over 1m science interested individuals and help us to provide you with insightful content whilst keeping our service free.

or

While we only use edited and approved content for Azthena answers, it may on occasions provide incorrect responses. Please confirm any data provided with the related suppliers or authors. We do not provide medical advice, if you search for medical information you must always consult a medical professional before acting on any information provided.

Your questions, but not your email details will be shared with OpenAI and retained for 30 days in accordance with their privacy principles.

Please do not ask questions that use sensitive or confidential information.

Read the full Terms & Conditions.