Real-time instance segmentation of spine endoscopy images using a YOLO-V11 deep convolutional neural network

BACKGROUND AND OBJECTIVES: Spine endoscopic surgery requires precise anatomical recognition and real-time bleeding detection to enhance surgical safety and efficiency. However, most existing computer vision studies do not offer real-time multi-class segmentation of the entire scene. This study aims to develop a real-time instance segmentation model based on YOLO-V11, capable of automatically identifying surgical instruments, critical anatomical structures, and bleeding sites in spine endoscopy. METHODS: The model was trained and evaluated using 2,181 annotated frames (training 1,741 / validation 216 / held-out test 224) from spine endoscopy images, covering nine key categories: instruments, bone, ligamentum flavum, dura, soft tissue, vessels, fat, the radiofrequency ablation wand, and bleeding focus. The segmentation model's performance was evaluated across various YOLO-V11 variants. RESULTS: Across five random seeds per variant on the same NVIDIA A100 GPU, YOLO-V11 l achieved the highest mask mAP50, defined as mean average precision at an intersection-over-union threshold of 0.50, with a value of 0.7735 ± 0.0030 (mean ± SD, n = 5). It also maintained real-time inference, reaching 270 ± 35 frames per second (FPS) with a per-frame latency of 3.75 ± 0.51 ms. All five variants exceeded 30 FPS, the widely accepted real-time threshold for surgical imaging. The model demonstrated high precision in segmenting vital structures such as the dura and ligamentum flavum and accurately identified bleeding regions. CONCLUSION: This research contributes to AI-driven surgical guidance systems, with potential applications in robotic-assisted spine surgery. The developed model enables real-time navigation, improving intraoperative decision-making and enhancing surgical outcomes. In particular, our YOLO-V11-based framework extends prior single-structure or lower-frame-rate multi-element approaches by delivering real-time multi-class instance segmentation of instruments, anatomical structures, and bleeding sites in spine endoscopy, with explicitly quantified inference latency.

Authors

Institutions

Publication Details

Journal
PLoS ONE
Published
2026-11-18
DOI
https://doi.org/10.1371/journal.pone.0361612
Primary Topic
Advanced Neural Network Applications
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Real-time instance segmentation of spine endoscopy images using a YOLO-V11 deep convolutional neural network

Jiho Kang, Taeshin Kim, Y. K. KIM C. S. LIM, Wonchang Choi et al.
3 links
PLoS ONE
Advanced Neural Network Applications
article

Real-time instance segmentation of spine endoscopy images using a YOLO-V11 deep convolutional neural network

Jiho Kang, Taeshin Kim, Y. K. KIM C. S. LIM, Wonchang Choi, Hyunjae Kim, Chulmin Lee, Noah Euyhyun Nam, Sang Kyeong Lee, Hyun Ji Jung, Gunyul Kim, Junsu Lee
article en

Abstract

BACKGROUND AND OBJECTIVES: Spine endoscopic surgery requires precise anatomical recognition and real-time bleeding detection to enhance surgical safety and efficiency. However, most existing computer vision studies do not offer real-time multi-class segmentation of the entire scene. This study aims to develop a real-time instance segmentation model based on YOLO-V11, capable of automatically identifying surgical instruments, critical anatomical structures, and bleeding sites in spine endoscopy. METHODS: The model was trained and evaluated using 2,181 annotated frames (training 1,741 / validation 216 / held-out test 224) from spine endoscopy images, covering nine key categories: instruments, bone, ligamentum flavum, dura, soft tissue, vessels, fat, the radiofrequency ablation wand, and bleeding focus. The segmentation model's performance was evaluated across various YOLO-V11 variants. RESULTS: Across five random seeds per variant on the same NVIDIA A100 GPU, YOLO-V11 l achieved the highest mask mAP50, defined as mean average precision at an intersection-over-union threshold of 0.50, with a value of 0.7735 ± 0.0030 (mean ± SD, n = 5). It also maintained real-time inference, reaching 270 ± 35 frames per second (FPS) with a per-frame latency of 3.75 ± 0.51 ms. All five variants exceeded 30 FPS, the widely accepted real-time threshold for surgical imaging. The model demonstrated high precision in segmenting vital structures such as the dura and ligamentum flavum and accurately identified bleeding regions. CONCLUSION: This research contributes to AI-driven surgical guidance systems, with potential applications in robotic-assisted spine surgery. The developed model enables real-time navigation, improving intraoperative decision-making and enhancing surgical outcomes. In particular, our YOLO-V11-based framework extends prior single-structure or lower-frame-rate multi-element approaches by delivering real-time multi-class instance segmentation of instruments, anatomical structures, and bleeding sites in spine endoscopy, with explicitly quantified inference latency.

PLoS ONEVol. 21(11)
Sogang University (KR), Konyang University (KR), Korea University (KR), Soonchunhyang University (KR), Soonchunhyang University Hospital Cheonan (KR), Samsung Electronics (South Korea) (KR), Pusan National University (KR), Chung-Ang University (KR)
Good health and well-being
Advanced Neural Network Applications
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon