aceAI detection of pneumothorax on supine neonatal chest radiographs in NICU patients
Source article: Development and evaluation of a deep learning model for computer-aided diagnosis of neonatal pneumothorax on chest radiographs
Abstract: Background Neonatal pneumothorax can progress rapidly, yet detection on supine chest radiographs remains challenging. Objective To develop and evaluate an artificial intelligence model for detecting pneumothorax on supine neonatal chest radiographs. Materials and methods This retrospective single-center study included neonates admitted to the neonatal intensive care unit between January 2011 and December 2024. The dataset comprised 648 radiographs from 288 neonates with pneumothorax and 5,511 radiographs from 3,…
Contested: both sides are scored from claims and sources, not community votes.
Hospital Universitari Doctor Peset, València 05 by 19Tarrestnom65. CC BY-SA 4.0 · https://creativecommons.org/licenses/by-sa/4.0
Researchers developed and evaluated a ResNet-18-based deep learning model to detect pneumothorax on supine neonatal chest radiographs using a retrospective dataset from a single NICU between 2011 and 2024. The model was pre-trained on normal adult radiographs and tested on held-out neonatal cases with patient-level splitting and a fixed threshold from cross-validation.
By the September 2026 publication date, the model had demonstrated high test-set discrimination with AUC 0.975 and strong negative predictive value, but with lower left-lung localization and potential sensitivity to underlying lung disease. The findings suggest promise for computer-aided diagnosis in a challenging supine imaging context, while generalizability remains uncertain without multi-center prospective validation.
- Retrospective single-center study of neonates admitted to the NICU between January 2011 and December 2024 with 648 pneumothorax radiographs from 288 neonates and 5,511 control radiographs from 3,377 controls.
- Model used publicly available normal adult chest radiographs for pre-training and knowledge distillation, with data split at the patient level and eight-fold cross-validation to set threshold.
- Grad-CAM localization matched clinically determined side in 84 of 92 right-lung instances (91.3%) but only 48 of 71 left-lung instances (67.6%).
A ResNet-18-based model trained on neonatal ICU radiographs detected pneumothorax on supine chest radiographs with AUC 0.975, 87.6% sensitivity and 95.3% specificity in the test set.
Model performance may be influenced by underlying pulmonary abnormalities, and lung-level localization was substantially lower for left-lung pneumothorax at 67.6% compared to right-lung.
The rundown
The study assembled 648 radiographs from 288 neonates with pneumothorax and 5,511 radiographs from 3,377 controls from a single NICU over 13 years, splitting data at the patient level into training, validation and test sets.
A ResNet-18-based model was pre-trained on publicly available normal adult chest radiographs with knowledge distillation, and Grad-CAM was used for lung-level localization while logistic regression identified factors associated with misclassification.
Test set results showed high overall discrimination but asymmetric localization accuracy by side, and the authors flagged pulmonary comorbidities as a factor linked to misclassification requiring further external validation.
Single-center retrospective design limits generalizability, and underlying pulmonary abnormalities may degrade performance; authors state external validation is needed before clinical use.
Sources
- Peer-reviewedPediatric Radiology2026-09-12
How should this claim be treated?
The debate