Canonical text (the exact bytes fingerprinted)

TRUVACE RECORD VERSION
record: TRV-2026-1031
version: 1
kind: certified
reason: Certified into the record
timestamp: 2026-09-09T06:06:31.071469Z
status: published
lens: g_space
sector: health
headline: Automated Diagnosis of Infantile Skull Fractures From X-Ray Images Using an Ensemble Deep Learning Model
dek: Background To develop an artificial intelligence (AI)-assisted model for detecting skull fractures in neonates and infants using plain radiographs, enhancing diagnostic accuracy while minimizing radiation exposure. Methods A retrospective dataset of skull X-rays from 1,184 patients with head trauma (2010-2021) was collected. Images underwent preprocessing, including background removal, Gaussian blurring, binarization, and CLAHE-based contrast enhancement. Three convolutional neural network architectures (ResNet-…
gain_title: An ensemble deep learning model combining AP and lateral skull X-rays detected skull fractures in neonates and infants with 91.6% accuracy and 0.938 AUC on external validation, improving diagnostic accuracy while reducing need for CT.
problem_title: (none)
trace_subject: (none)
gain_reading: An ensemble deep learning model combining AP and lateral skull X-rays detected skull fractures in neonates and infants with 91.6% accuracy and 0.938 AUC on external validation, improving diagnostic accuracy while reducing need for CT.
gain_evidence: achieved an overall AUC of 0.938, with a classification accuracy of 91.6% on the external validation dataset
problem_reading: (none)
problem_evidence: (none)
quick_read: By September 7, 2026, researchers reported developing an AI-assisted ensemble model to detect skull fractures in neonates and infants from plain radiographs. Using a retrospective set of 1,184 patients from 2010-2021, they preprocessed images with CLAHE and trained three CNNs, with DenseNet-121 performing best, then combined AP and lateral views into an ensemble that achieved 0.938 AUC and 91.6% accuracy on an external set of 460 images.

The result matters for pediatric emergency care because plain X-rays are lower radiation than CT but fractures are hard to distinguish from normal sutures in infants. The model is presented as improving clinical decision-making and reducing unnecessary CT scans, but the supplied text shows only retrospective validation with internal and external image sets, leaving prospective clinical impact, workflow integration, and generalizability beyond the studied datasets uncertain.
limitation: 
tag: Evidence-backed gain
key_points: Retrospective dataset of 1,184 patients with head trauma from 2010-2021 used for training. | Preprocessing included background removal, Gaussian blurring, binarization, and CLAHE-based contrast enhancement. | Three CNN architectures tested: ResNet-50, DenseNet-121, EfficientNet-B5, with DenseNet-121 optimized using CLAHE performing best. | Ensemble combined anterior-posterior and lateral views and was evaluated on internal 4,298 images and external 460 images. | Expert annotation by neurosurgeons and radiologists helped differentiate fractures from cranial sutures, reducing false positives.
rundown: The team collected skull X-rays from 1,184 patients with head trauma between 2010 and 2021 and applied preprocessing steps including background removal, Gaussian blurring, binarization, and CLAHE-based contrast enhancement. They trained ResNet-50, DenseNet-121, and EfficientNet-B5, finding DenseNet-121 with CLAHE achieved mean AUC 0.926 AP and 0.911 lateral, then built an ensemble combining both views.

Performance was measured on 4,298 internal images and 460 external images, with the ensemble reaching overall AUC 0.938 and 91.6% accuracy externally. The authors report expert annotation by neurosurgeons and radiologists improved reliability by enabling differentiation between fractures and normal structures such as cranial sutures.
sources:
- peer_reviewed | Journal of Korean Medical Science | https://doi.org/10.3346/jkms.2026.41.e232 | 2026-09-07
prev: 0000000000000000000000000000000000000000000000000000000000000000

sha256

46bae60096fd3e42567cdcaf0ef9081afe769fefe580825a368522e189588a27

previous

0000000000000000000000000000000000000000000000000000000000000000