For critically ill children who depend on mechanical ventilation, the position of a thin plastic tube just a few millimeters too deep or too shallow can mean the difference between stable breathing and a life-threatening complication. Now, a team of researchers at Cincinnati Children’s Hospital Medical Center has demonstrated that artificial intelligence can reliably detect and locate endotracheal tubes on pediatric chest X-rays, a task that consumes enormous radiologist time and where errors are disturbingly common in the youngest patients. The study, published in Pediatric Radiology, describes a two-stage deep learning pipeline that achieved near-perfect accuracy in detecting whether a breathing tube is present and promising precision in pinpointing exactly where its tip sits within a child’s airway.
The clinical stakes are considerable. Endotracheal tubes, or ETTs, are inserted through the mouth or nose into the trachea to deliver mechanical ventilation in intensive care units and operating rooms. When properly positioned, the distal tip of the tube should terminate between the thoracic inlet and the carina, the ridge where the trachea divides into the two main bronchi. But malposition is strikingly frequent in children, with reported incidences reaching up to 35 percent in infants younger than one year and more than 10 percent in children up to age ten. A tube inserted too far can enter one bronchus and ventilate only a single lung, causing hypoxemia, pneumothorax, or even death; a tube placed too high risks accidental extubation and vocal cord injury.
Because these risks demand constant vigilance, chest radiography remains the standard method for confirming tube placement after intubation and during follow-up, generating a relentless stream of images that radiologists must interpret alongside everything else in pediatric intensive care and emergency settings. Artificial intelligence has already made meaningful inroads on this problem for adults, with multiple deep learning systems demonstrating the ability to detect tubes, segment their course, and measure the distance from the tube tip to the carina. What works in adults, however, does not automatically translate to children. Pediatric chest radiographs are far more heterogeneous, shaped by dramatic variation in body size, lung and skeletal development, and imaging technique across the age spectrum from premature neonates to teenagers.
There was another, more subtle flaw in much of the existing research that the Cincinnati team set out to fix. Most prior studies trained and evaluated their segmentation algorithms only on images already known to contain a tube. That simplification makes model development easier but bears little resemblance to clinical reality, where radiologists interpret every radiograph without knowing in advance whether a device is present. Applied to images without tubes, standalone segmentation models tend to hallucinate, mistaking tracheostomy tubes, feeding tubes, and other radiopaque lines for endotracheal tubes. Such false-positive segmentations go unnoticed by conventional localization metrics, which are typically computed only on images where the tube was correctly found.
To build a system suited to the messiness of real practice, the researchers assembled a cohort of 1,000 pediatric chest radiographs drawn from 28,188 single-view studies performed at their institution during 2021, representing 553 patients younger than 18. The cohort was deliberately balanced, comprising 476 tube-positive and 524 tube-negative images, with the negative cases including other indwelling devices such as tracheostomy and enteric tubes as well as images free of any lines. After re-review, 24 images initially tagged as tube-positive were reassigned. Three trained data analysts, supervised by a pediatric radiologist with 18 years of experience, manually created pixel-level segmentation masks of the full tube course using the open-source 3D Slicer platform, and distal tip coordinates were independently annotated by two pediatric radiologists whose disagreement, quantified as a mean absolute error of 2.01 millimeters on the test set, served as a human benchmark.
The pipeline itself operates in two stages. Every image is first zero-padded to a square, resized to 1,024 by 1,024 pixels, and enhanced with contrast-limited adaptive histogram equalization, a technique that sharpens local contrast in radiographic structures. A stage-one classification model then decides whether a tube is present at all. The team benchmarked five backbone architectures, including Inception, ResNet, DenseNet, MobileNet, and a Vision Transformer, each initialized with ImageNet pre-trained weights, and tested both frozen feature extraction and full fine-tuning across 20 configurations. A ResNet classifier with a 128-node fully connected head, trained with fine-tuning, won out on validation performance. If the classifier reports no tube, the pipeline stops; otherwise the image passes to a stage-two U-Net segmentation model, selected from nine architectural variants, that traces the entire course of the tube. Postprocessing keeps only the largest connected component of the predicted mask, and a classic thinning algorithm detects the mask’s endpoint to yield the distal tip coordinates.
On a held-out test set of 220 images the pipeline’s results were striking. Detection accuracy reached 97.7 percent, with an area under the receiver operating characteristic curve of 0.994, and performance was statistically indistinguishable between younger and older children and between boys and girls, suggesting the model carries no obvious age- or sex-related bias. Grad-CAM visualization, a technique that highlights which image regions drive a neural network’s decisions, showed the classifier attending to anatomically meaningful structures such as the upper trachea and carina rather than spurious features. For localization, the pipeline achieved a Dice similarity coefficient of 0.74 for mask overlap and a mean tip localization error of 6.59 millimeters. Notably, incorporating the classification stage slashed false-positive segmentations on tube-negative images from 17 to just 3 compared with the standalone segmentation model, at the cost of only two false negatives.
The error analysis reveals both the promise and the remaining gaps. In the single false-negative case, the classifier assigned a tube probability of just 0.06, below the optimized threshold of 0.2, because the tube sat unusually high with only a short segment visible. In a representative false positive, the model mistook an enteric tube and an esophageal probe for a breathing tube, a reminder that other tubular devices remain the chief source of confusion. Bland-Altman analysis showed that the model estimates the tube tip’s lateral position on the image with tight precision, but its vertical estimate, which corresponds most closely to the clinically critical depth of the tube, carried far wider limits of agreement. The authors caution that image-based measurements also depend on patient positioning and tracheal orientation and cannot be read directly as anatomical depth.
Honest benchmarking against human performance tempers the enthusiasm. The radiologists disagreed with each other by only about 2 millimeters, while the AI’s 6.59 millimeter average error remains well above that human-level agreement, a gap that matters most in the smallest neonates, where a few millimeters can separate a safe tube position from a dangerous one. The study also has other limits: it is a single-center retrospective pilot, the cohort skewed toward younger children, and the pipeline detects and localizes tubes without yet judging whether the position is appropriate relative to landmarks like the carina, a capability the team says could come from adding automated carina detection in future work. Still, by proving that a two-stage design can tame the false-positive problem in mixed clinical populations, the Cincinnati team has laid a credible foundation for AI systems that could one day triage post-intubation films in pediatric ICUs worldwide, flagging misplaced tubes for urgent human review before complications ever develop.
The design choices behind the study reflect deliberate responses to known pitfalls in medical imaging AI. The balanced target of roughly 500 tube-positive and 500 tube-negative images was chosen to minimize class imbalance during development, and the sample size was informed by an earlier adult deep learning study that succeeded with only 292 tube-positive radiographs. The Cincinnati team reasoned that the greater anatomical variability of children warranted a substantially larger positive sample, illustrating how pediatric AI research often requires more data than comparable adult work to reach similar reliability.
The annotation workflow also merits attention as a model for scalable ground-truth creation. Rather than relying solely on scarce radiologist time, the investigators trained three data analysts with two to three years of experience to draw pixel-level masks in 3D Slicer, after which a supervising pediatric radiologist with 18 years of experience reviewed and corrected every mask. Distal tip coordinates were annotated independently by two radiologists, and the average of their coordinates served as the evaluation reference. This tiered approach, in which analysts perform labor-intensive tracing while physicians verify, offers a practical template for other institutions seeking to build annotated pediatric imaging datasets without exhausting specialist capacity.
Image preparation played a quiet but important role in performance. All radiographs were zero-padded along the shorter dimension to preserve aspect ratio, resized to 1,024 by 1,024 pixels, and processed with contrast-limited adaptive histogram equalization, which enhances local contrast in ways that help neural networks distinguish thin radiopaque devices from surrounding soft tissue and bone. Such preprocessing steps are often decisive in radiograph analysis, where tube visibility varies with patient size and exposure technique.
The open-access publication lowers barriers for other pediatric centers to reproduce or extend the work. Because the pipeline separates detection from localization, future modules such as automated carina detection or position classification could be layered onto the existing framework, moving the field closer to fully automated assessment of tube placement in children.
Subject of Research: Automated detection and localization of endotracheal tubes on pediatric chest radiographs using a two-stage deep learning pipeline
Article Title: A two-stage deep learning pipeline for automated detection and localization of endotracheal tubes on pediatric chest radiographs: a pilot study
Article References: Li, H., Zhang, B., Somasundaram, E., Shabanian, M., Taylor, Z., Mahalingam, N., Lu, Z., Zhang, B., Standage, S. W., Schooler, G. R., He, L., & Towbin, A. J. (2026). A two-stage deep learning pipeline for automated detection and localization of endotracheal tubes on pediatric chest radiographs: a pilot study. Pediatric Radiology. https://doi.org/10.1007/s00247-026-06767-z
Image Credits: AI Generated
DOI: 10.1007/s00247-026-06767-z
Keywords: artificial intelligence, deep learning, pediatric radiology, endotracheal tube, chest X-ray, image segmentation, U-Net, ResNet, pediatric intensive care, computer vision, medical imaging, tube localization
Cite Scienmag News
APA
MLA
Chicago
Blake Davidson. (September 12, 2026). AI Learns to Spot Life-Saving Breathing Tubes on Children’s Chest X-Rays. Scienmag. https://scienmag.com/ai-learns-to-spot-life-saving-breathing-tubes-on-childrens-chest-x-rays/
Blake Davidson. “AI Learns to Spot Life-Saving Breathing Tubes on Children’s Chest X-Rays.” Scienmag, 12 September 2026, https://scienmag.com/ai-learns-to-spot-life-saving-breathing-tubes-on-childrens-chest-x-rays/. Accessed 12 September 2026.
Blake Davidson. “AI Learns to Spot Life-Saving Breathing Tubes on Children’s Chest X-Rays.” Scienmag. September 12, 2026. https://scienmag.com/ai-learns-to-spot-life-saving-breathing-tubes-on-childrens-chest-x-rays/
Copy citation
Download RIS
Tags: AI in critical care airway managementAI-assisted diagnosis in pediatric radiologyArtificial Intelligenceautomated detection of endotracheal tubes in childrenchallenges of tube placement in pediatric patientschest X-raycomputer visiondeep learningdeep learning for endotracheal tube detectiondeep learning pipelines for medical imagingendotracheal tubeimage segmentationlife-saving airway management in childrenmachine learning for medical image analysisMedical Imagingpediatric chest X-ray analysis using artificial intelligencepediatric intensive carepediatric radiologypediatric ventilator tube placement accuracyradiologist support for pediatric intensive careResNetsafety and accuracy of pediatric intubationtube localizationU-Net
