Real-Time Glottis Detection Framework via Spatial-decoupled Feature Learning for Nasal Transnasal Intubation

Jinyu Liu; Gaoyang Zhang; Yang Zhou; Ruoyi Hao; Yang Zhang; Hongliang Ren

cs.CVMar 8, 2026

Real-Time Glottis Detection Framework via Spatial-decoupled Feature Learning for Nasal Transnasal Intubation

Authors: Jinyu Liu, Gaoyang Zhang, Yang Zhou, Ruoyi Hao, Yang Zhang, Hongliang Ren

Organizations: Hubei Key Laboratory of Modern Manufacturing Quality Engineering, Hubei University of Technology, Nanhu Avenue 28, Wuhan, 430068, Hubei, China · School of Mechanical Science and Engineering, Huazhong University of Science and Technology, Luoyu Road 1037, Wuhan, 430074, Hubei, China · Department of Electronic Engineering, The Chinese University of Hong Kong, Shatin, Hong Kong, 999077, China · Key Laboratory of Symbolic Computation and Knowledge Engineering, Ministry of Education, Jilin University, Changchun, 130012, Jilin, China ·

Abstract

Nasotracheal intubation (NTI) is a vital procedure in emergency airway management, where rapid and accurate glottis detection is essential to ensure patient safety. However, existing machine assisted visual detection systems often rely on high performance computational resources and suffer from significant inference delays, which limits their applicability in time critical and resource constrained scenarios. To overcome these limitations, we propose Mobile GlottisNet, a lightweight and efficient glottis detection framework designed for real time inference on embedded and edge devices. The model incorporates structural awareness and spatial alignment mechanisms, enabling robust glottis localization under complex anatomical and visual conditions. We implement a hierarchical dynamic thresholding strategy to enhance sample assignment, and introduce an adaptive feature decoupling module based on deformable convolution to support dynamic spatial reconstruction. A cross layer dynamic weighting scheme further facilitates the fusion of semantic and detail features across multiple scales. Experimental results demonstrate that the model, with a size of only 5MB on both our PID dataset and Clinical datasets, achieves inference speeds of over 62 FPS on devices and 33 FPS on edge platforms, showing great potential in the application of emergency NTI.

Figures & tables

Fig. 1 : The comparative visualization with state-of-the-art methods on the PID dataset is illustrated, where the vertical axis represents the inference speed (FPS) measured on terminal devices. The bubble chart shows that our method achieves a favorable balance between AP50 and FPS while maintaining a compact model size. The area of each bubble corresponds to the model size.

Fig. 2 : Overview of the proposed Mobile GlottisNet. The input image is first processed by the backbone network to generate multi-scale feature maps, which are then fused through the FPN to produce enhanced multi-scale feature representations. Subsequently, the classification and regression tasks are decoupled, and dynamic label assignment strategies are applied to adjust anchor point allocation. The mean IoU of candidate boxes is calculated to set an adaptive threshold. Finally, convolutional kernel sampling positions are adjusted based on the Offset Field in the adaptive feature decoupling module, and the final detection results are derived from class scores and bounding box offsets.

Require : - Predicted bounding boxes

\{b_{i}\}

across

L

FPN scales,

- Ground-truth class

C_{j}

for the

j

-th glottis,

- Hyperparameters: balance factor

\lambda

, top-

K

count.

Ensure : Selected candidate set

T_{j}

for training classification/regression.

1. Cost Matrix Calculation

For each predicted box

b_{i}

and ground-truth

j

C_{ij}=-\lambda\cdot\log\left(\frac{e^{P_{i}^{C_{j}}}}{\sum_{k=1}^{K}e^{P_{i}^{k}}}\right)+(1-\lambda)\cdot\left|1-\text{IoU}(b_{i},b_{j})\right|

Algorithm 1 Glottis Candidate Selection with Dynamic Thresholding

Fig. 3 : Attention heatmaps throughout the network. Feature disentanglement progressively refines spatial attention, culminating in accurate focus on the anatomical glottic aperture.

Input : Feature map

F\in\mathbb{R}^{H\times W\times C}

, regular grid

\mathcal{R}

, convolutional kernel weights

W\in\mathbb{R}^{k\times k\times C}

;

Parameters: Offset generation network parameters

\theta

, uncertainty ratio

\eta

;

1. Offset Generation:

Compute offsets

\Delta p^{\text{cls/reg}}=\Phi(F;\theta)

, where for all

q\in\mathcal{R}

\Delta p_{q}^{\text{cls/reg}}=(\Delta x_{q}^{\text{cls/reg}},\Delta y_{q}^{\text{cls/reg}})

;

2. Deformable Sampling:

For each output position

p

F_{\text{cls}}(p)=\sum_{q\in\mathcal{R}}W_{\text{cls}}(q)\cdot F\bigl(p+q+\Delta p_{q}^{\text{cls}}\bigr)

, and

F_{\text{reg}}(p)=\sum_{q\in\mathcal{R}}W_{\text{reg}}(q)\cdot F\bigl(p+q+\Delta p_{q}^{\text{reg}}\bigr)

;

3. Gradient Backpropagation Mechanism:

Algorithm 2 Deformable Convolution Forward Computation Pipeline

Fig. 4 : The nasotracheal intubation robot system is an advanced medical system that improves intubation accuracy and minimizes human error. It utilizes a fiberoptic bronchoscope with a bendable robotic arm, dynamically adjusting insertion paths through real-time feedback control.

Fig. 5 : Sample images from the Glottis dataset. The dataset includes high-speed videoendoscopic recordings from both healthy individuals and patients with laryngeal disorders, collected across different clinical settings and imaging devices.

Panel of Figure 8 — Fig. 6 : During NTI, the glottis undergoes substantial scale variation under complex and low-light conditions, and the surrounding anatomy is intricate. These factors often degrade the accuracy of conventional detectors.

Methods	Backbone	PID (%)			Clinical (%)			Model size (MB)	Infer (FPS)
Methods	Backbone	mAP	$AP_{50}$	$AP_{75}$	mAP	$AP_{50}$	$AP_{75}$	Model size (MB)	Device	Edge
ATSS Zhang et al. [2020]	ResNet50	30.3	59.5	28.4	17.8	39.5	13.6	251	21.6	12.1
CenterNet Zhou et al. [2019]	ResNet18	25.5	54.9	24.0	12.6	25.9	10.9	114	50.0	25.3
CentripetalNet Zhou et al. [2020]	Hourglass104	28.0	49.2	26.5	5.3	9.9	4.7	2416	11.2	5.9
Faster RCNN Ren et al. [2017]	ResNet50	29.1	55.1	27.6	32.4	61.0	29.3	323	51.3	26.1
GFL Li et al. [2020]	ResNet101	33.1	61.3	32.7	34.9	61.5	35.7	401	30.1	16.4

Table 1: Comparison of accuracy with the state-of-the-art methods on the PID and clinical datasets.

Methods	Backbone	Glottis (%)			Model size (MB)	Inference (FPS)
Methods	Backbone	$mAP$	$AP_{50}$	$AP_{75}$	Model size (MB)	Inference (FPS)
ATSS Zhang et al. [2020]	ResNet50	53.7	84.8	60.1	251	21.6
BoxInst Tian et al. [2021]	ResNet101	53.3	82.7	58.5	447	45.7
CondInst Tian et al. [2020]	ResNet50	41.1	79.1	36.9	279	46.8
CMask RCNN Cai and Vasconcelos [2021]	ResNet50	55.9	83.2	62.9	201	16.4
Conditional DETR Meng et al. [2021]	ResNet50	7.1	21.9	3.0	512	34.4

Table 2: Comparison of accuracy with the state-of-the-art methods on the Glottis dataset.

Backbone	mAP	$AP_{50}$	$AP_{75}$	Inference (FPS)	Model Size (MB)
ResNet101	38.8	72.5	35.4	26.6	205
ResNet50	33.3	66.3	30.6	31.5	129
ResNet18	30.1	61.6	29.2	40.2	77
ShuffleNet	18.7	42.4	11.6	65.4	73
Swin-T	35.3	63.7	28.6	21.2	285
EfficientNet	28.7	54.4	21.9	51.4	92

Table 3: Performance of various backbones on the PID dataset.

TopK	$mAP$	$AP_{50}$	$AP_{75}$	TopK	$mAP$	$AP_{50}$	$AP_{75}$
1	27.4	51.0	26.7	11	33.5	63.0	32.5
3	32.3	57.0	31.5	13	30.6	58.6	30.0
5	31.7	57.5	31.8	15	30.2	57.2	27.9
7	35.0	63.4	34.6	17	32.0	59.6	31.1
9	33.3	66.3	30.6	19	35.7	66.0	35.5

Table 4: Comparison of TopK in dynamic thresholding & hierarchical sample allocation on PID dataset.

Channels	$mAP$	$AP_{50}$	$AP_{75}$	Infer (FPS)	Model Size (MB)
256	33.3	66.3	30.6	31.5	129.1
128	32.4	68.2	26.8	36.4	104.1
64	32.7	66.6	27.9	43.9	97.3
32	28.2	57.1	25.1	48.6	95.3
16	25.7	54.0	20.8	50.5	94.7
8	20.5	45.7	16.5	51.7	94.5

Table 5: Performance of Mobile GlotissNet with MobileNetv3 in various channels on PID dataset.

Panel of Figure 14 — Fig. 7 : Qualitative comparison among ground-truth labels, model predictions, and video detection results across glottis, clinical, and phantom datasets.

Efficient Object Detection

Efficient object detection reduces the computation, latency, memory, and energy needed to find and locate objects in images and video.

See all 95 papers

A Real-time Scale-robust Network for Glottis Segmentation in Nasal Transnasal Intubation
Apr 30, 2026Yang Zhou, Chaoyong Zhang, Ruoyi Hao +3Medical Image Segmentation Semantic Segmentation
School of Mechanical Science & Engineering Huazhong University of Science and Technology, Wuhan 430074, China · Department of Electronic Engineering, The Chinese University of Hong Kong, Hong Kong, China · Shun Hing Institute of Advanced Engineering, The Chinese University of Hong Kong, Hong Kong 999077, China +5
A Detection-Gated Pipeline for Robust Glottal Area Waveform Extraction and Clinical Pathology Assessment
Mar 2, 2026Harikrishnan Unnikrishnan, Rita PatelMedical Image Segmentation Endoscopy
¹Orchard Robotics, San Francisco, California, USA. · ²Department of Speech, Language, and Hearing Sciences, Indiana University, Bloomington, Indiana, USA. · ³Department of Otolaryngology–Head and Neck Surgery, Indiana University School of Medicine, Indianapolis, Indiana, USA.
RT-DETRv4: Painlessly Furthering Real-Time Object Detection with Vision Foundation Models
Oct 29, 2025Zijun Liao, Yian Zhao, Xin Shan +5Real-Time Object Detection Efficient Object Detection
School of Electronic and Computer Engineering, Peking University, Shenzhen, China · Department of Automation and BNRist, Tsinghua University, Beijing, China

Apr 30, 2026J/K move · Enter open · S save

A Real-time Scale-robust Network for Glottis Segmentation in Nasal Transnasal Intubation

Yang Zhou, Chaoyong Zhang, Ruoyi Hao, Huilin Pan+2

School of Mechanical Science & Engineering Huazhong University of Science and Technology, Wuhan 430074, China · Department of Electronic Engineering, The Chinese University of Hong Kong, Hong Kong, China · Shun Hing Institute of Advanced Engineering, The Chinese University of Hong Kong, Hong Kong 999077, China+5

Medical Image Segmentation Semantic Segmentation

Open paper AskPDF arXiv

Nasotracheal intubation (NTI) is a critical clinical procedure for establishing and maintaining patient airway patency. Machine-assisted NTI has emerged as a pivotal approach for optimizing procedural efficiency and minimizing manual intervention. However, visual detection algorithms employed for NTI navigation encounter significant challenges, including complex anatomical environments and suboptimal illumination conditions surrounding the glottis. Additionally, the glottis presents considerable scale variability throughout the procedure, initially appearing as a small, difficult-to-capture structure before expanding to occupy nearly the entire field of view. Moreover, traditional visual detection methods often have high computational costs, making real-time, high-precision detection on portable devices challenging. To enhance NTI efficacy and address these challenges, this paper proposes a novel glottis segmentation framework optimized for vision-assisted NTI applications. First, we designed a lightweight, multi-receptive field feature extraction module to reduce intra-class differences, achieving robustness to scale variations of the glottis. This module was then stacked to form the backbone and neck of our network. Subsequently, we developed an advanced label assignment method and redefined the number of samples to further reduce intra-class differences and enhance accuracy in the complex NTI environment. Experiments on three distinct datasets demonstrate that our network surpasses state-of-the-art algorithms, achieving a segmentation mDice of 92.9% with a compact model size of 19 MB and an inference speed exceeding 170 frames per second. % Our code and datasets will be open-sourced on GitHub after the manuscript is accepted. Our code and datasets are available at https://github.com/HBUT-CV/GlottisNet.

Abstract

Efficient Object Detection

Explore similar work

A Real-time Scale-robust Network for Glottis Segmentation in Nasal Transnasal Intubation

A Detection-Gated Pipeline for Robust Glottal Area Waveform Extraction and Clinical Pathology Assessment

RT-DETRv4: Painlessly Furthering Real-Time Object Detection with Vision Foundation Models