An Artificial Neural Network for Nasogastric Tube Position Decision Support
2026-01-24https://doi.org/10.1148/atlas.1769272088336
30
Overview
Schema Version
https://atlas.rsna.org/schemas/2025-11/model.json
Name
An Artificial Neural Network for Nasogastric Tube Position Decision Support
Link
https://www.ncbi.nlm.nih.gov/pmc/articles/PMC10077078/
Indexing
Keywords: nasogastric tube, NGT malposition, chest radiograph, decision support, feed/do not feed, bronchial insertion, emergency department
Content: CH, ER, RS
RadLex: RID5557, RID5566, RID43583, RID43594, RID4897
Author(s)
Ignat Drozdov
Rachael Dixon
Benjamin Szubert
Jessica Dunn
Darren Green
Nicola Hall
Arman Shirandami
Sofia Rosas
Ryan Grech
Srikanth Puttagunta
Mark Hall
David J. Lowe
Organization(s)
Bering Limited
Queen Elizabeth University Hospital, NHS Greater Glasgow and Clyde
Department of Radiology, Queen Elizabeth University Hospital, Glasgow
Institute of Health and Wellbeing, University of Glasgow
NHS Greater Glasgow and Clyde Safe Haven
Industrial Centre for AI Research in Digital Diagnostics (iCAIRD)
Canon Medical Research Europe Limited
Version
1.0
License
Text: CC BY 4.0
URL: https://creativecommons.org/licenses/by/4.0/
Contact
moc.hcraesergnireb@vodzordi
Funding
Supported by Bering Limited and the Industrial Centre for AI Research in Digital Diagnostics (iCAIRD), funded by the UKRI/Innovate UK Industrial Strategy Challenge Fund (project no. 104690).
Ethical review
Delegated research ethics approval granted by the Local Privacy and Advisory Committee at NHS Greater Glasgow and Clyde (reference: 104690/WP6/S1). In Scotland, informed consent was not required for use of routinely collected patient data through an approved Safe Haven.
Date
Published: 2023-02-01
Created: 2022-08-09
References
[1] Drozdov I, Dixon R, Szubert B, et al.. "An Artificial Neural Network for Nasogastric Tube Position Decision Support". Radiology: Artificial Intelligence. 2023;5(2):e220165. 2023-02-01. doi:10.1148/ryai.220165. PMID: 37035435. PMCID: PMC10077078.
Model
Architecture
Ensemble of modified InceptionV3 convolutional neural networks operating at two input resolutions (764×764 and 1024×1024) with global average pooling, dense ReLU layer, dropout, and softmax output; pretrained (ImageNet) then pretrained on chest radiograph normal/abnormal task and fine-tuned for NGT position classification.
Clinical benefit
Assists junior physicians in determining safety of feeding in patients with nasogastric tubes by detecting NGT malposition on chest radiographs and improving agreement with radiologists.
Clinical workflow phase
Clinical decision support systems in emergency and acute care settings.
Degree of automation
Decision support; provides class probabilities to support clinician judgment (non-autonomous).
Indications for use
Detection of nasogastric tube malposition on adult frontal chest radiographs (AP/PA) in hospital settings (e.g., emergency department) to support feed/do not feed decisions.
Input
Adult frontal chest radiographs (AP/PA) in DICOM format; images resized to 764×764 or 1024×1024 for model input.
Instructions
Display AI-generated probabilities for NGT position classes (satisfactory, malpositioned, bronchial) above the chest radiograph; clinicians consider outputs alongside their own judgment to decide feed/do not feed.
Limitations
No tube segmentation or localization; evaluated retrospectively and not on radiologist workstations; class balance may not reflect real-world prevalence; lateral chest radiographs not included; misclassifications noted in cases such as gastric pull-up surgery and hiatus hernia; model performance sensitive to concept drift related to radiography system manufacturer, patient age, and acquisition department.
Output
CDEs: RDE1533.1, RDE1533.2
Description: Per-image class probabilities for NGT position: satisfactory, malpositioned, or bronchial; used to inform feed/do not feed decisions.
Recommendation
Use as a second-opinion decision support tool for junior physicians interpreting NGT position on chest radiographs.
Use
Intended: Decision support, Detection and diagnosis
Out-of-scope: Decision support
Excluded: Other
User
Intended: Referring provider, Radiologist, Referring physician
Out-of-scope: Patient, Layperson
Excluded: Other