Related Experiment Video
Updated: Jul 26, 2025
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
2.8K
Development of a Deep Learning Model for Retinal Hemorrhage Detection on Head Computed Tomography in Young Children
Fatma Gunturkun1, Berna Bakir-Batu2, Adeel Siddiqui3
1Quantitative Sciences Unit, Department of Medicine, Stanford University, Palo Alto, California.
JAMA Network Open
|June 22, 2023
Summary
Deep learning models can detect retinal hemorrhages (RH) on pediatric head CT scans, aiding in the diagnosis of abusive head trauma (AHT). This AI tool can help identify children needing further examination for AHT.
Area of Science:
- Medical imaging analysis
- Artificial intelligence in diagnostics
- Pediatric medicine
Background:
- Abusive head trauma (AHT) is a critical concern in child abuse cases.
- Retinal hemorrhages (RH) are a key indicator of AHT but are often missed.
- Current head CT scans have limited ability to detect RH.
Purpose of the Study:
- To evaluate the efficacy of deep learning (DL) models in detecting RH on pediatric head CT scans.
- To assess if AI can improve the identification of RH, a marker for AHT.
Main Methods:
- A diagnostic study analyzed head CT scans from 301 children with diagnosed AHT.
- A DL model was trained and tested on segmented globes with and without RH.
- Performance was compared against gradient boosting machine (GBM) models using clinical and CT data.
Main Results:
- The DL model achieved a sensitivity of 79.6% and specificity of 79.2% for RH detection.
- Combined DL and clinical data models showed higher specificity than clinical data alone.
- The DL model demonstrated an AUC of 0.83 for predicting RH presence.
Conclusions:
- Deep learning-based image analysis of pediatric head CTs can effectively predict RH.
- Integrating DL into CT software could enhance AHT diagnosis and guide clinical decisions.
- This AI tool may help prioritize fundoscopic examinations for at-risk children.
