Automated ventricular segmentation in pediatric hydrocephalus: how close are we?

Area of Science:

• Medical imaging analysis
• Artificial intelligence in healthcare
• Pediatric neuroimaging

Background:

• Advancements in imaging hardware and data availability have spurred progress in brain segmentation.
• Existing segmentation models often have limited or unknown performance in pediatric populations, particularly for conditions like hydrocephalus.

Purpose of the Study:

• To evaluate the accuracy of current automated ventricular segmentation tools (FastSurfer and QuickNAT) in pediatric hydrocephalus cases.
• To identify factors influencing the performance of these segmentation models in children.

Main Methods:

• Utilized 40 scans from 32 pediatric patients diagnosed with hydrocephalus.
• Compared automated segmentation outputs from FastSurfer and QuickNAT against manual segmentations using Dice Similarity Coefficient (DSC).

Main Results:

• Both FastSurfer and QuickNAT showed poor performance with an average DSC of 0.61.
• Younger age and larger ventricular volumes were significantly associated with lower segmentation accuracy (lower DSC).
• No significant difference in performance was observed between the two models, nor were factors like gender or MRI magnet strength significant.

Conclusions:

• Current automated segmentation models demonstrate significant performance gaps when applied to pediatric hydrocephalus.
• Future model development requires careful consideration of training data to address these limitations and enable clinical deployment.