Deep Learning Auto-Segmentation Network for Pediatric Computed Tomography Data Sets: Can We Extrapolate From Adults?

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiation Oncology

Background:

• AI-based auto-segmentation offers efficiency in organ contouring for radiation therapy.
• Performance of AI models on pediatric CT data and cross-scanner compatibility require investigation.

Purpose of the Study:

• Evaluate AI auto-segmentation models trained on adult data when applied to pediatric CT scans.
• Assess the performance improvement with the inclusion of pediatric training data.
• Examine the cross-scanner compatibility of these AI models.

Main Methods:

• Utilized the nnU-Net framework to train segmentation models on adult, pediatric, and combined CT datasets.
• Trained models on 290-300 cases per organ for 7 pelvic/thoracic organs.
• Evaluated performance using Dice Similarity Coefficients (DSC) on a database of 459 pediatric and 950 adult CT scans.

Main Results:

• AI models trained solely on adult data showed poor performance on pediatric scans (DSC < 0.5 for bladder, spleen in 0-2 age group).
• Incorporating pediatric data significantly improved performance across all age groups (mean DSC > 0.85).
• Consistent performance was observed for larger organs, and models showed robust cross-scanner generalization.

Conclusions:

• Pediatric data inclusion is crucial for optimal AI auto-segmentation across all age groups.
• The AI models exhibit strong cross-scanner generalization, supporting clinical use.
• Dataset diversity is vital for developing robust AI systems in medical imaging.