High-level feature-guided attention optimized neural network for neonatal lateral ventricular dilatation prediction

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neonatal Neurology

Background:

• Periventricular-intraventricular hemorrhage can lead to posthemorrhagic ventricular dilatation or hydrocephalus if not promptly diagnosed.
• Cranial ultrasound scans are standard for diagnosis, but manual image analysis has limitations.

Purpose of the Study:

• To develop a predictive model using modified inception (MI) and high-level feature-guided attention (HFA) modules for predicting neonatal lateral ventricular dilation from ultrasound images.

Main Methods:

• Utilized MI modules to reduce data size and HFA modules for semantic information extraction.
• Trained and tested the model on 710 neonates (1420 lateral ventricles) with data split 7:3.
• Evaluated performance across coronal and parasagittal planes, considering baseline characteristics and performing ablation analyses.

Main Results:

• The proposed approach achieved superior performance metrics, including a Youden index of 0.65 and an ROC AUC of 0.84.
• Demonstrated significant Cramer's V values (0.488 coronal, 0.713 parasagittal), indicating strong predictive power.
• Identified factors like male sex and singleton birth positively correlated with performance, while gestational age did not.

Conclusions:

• The developed attention-optimized algorithm enables rapid and accurate prediction of ventricular dilation.
• This novel approach outperforms traditional methods in validity across different imaging planes.
• Population baseline characteristics can influence the overall performance of such predictive algorithms.