Population-Driven Synthesis of Personalized Cranial Development From Cross-Sectional Pediatric CT Images

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Pediatric Development

Background:

• Accurate prediction of pediatric growth is vital for identifying developmental anomalies.
• Traditional methods often require scarce longitudinal data or have limited generalizability.
• Existing deep learning models struggle with major anatomical changes in childhood development.

Purpose of the Study:

• To present a novel deep learning method for image synthesis to predict pediatric development.
• To enable personalized predictions of child growth using only cross-sectional data.
• To overcome limitations of existing methods in modeling anatomical changes.

Main Methods:

• Developed a generative adversarial network (GAN) with a Siamese cyclic encoder-decoder architecture.
• Incorporated an identity preservation mechanism for learning age- and sex-independent representations.
• Trained the model using cross-sectional head CT images from 2,014 subjects (ages 0-10).

Main Results:

• The model demonstrated state-of-the-art performance in predicting pediatric development.
• Evaluation was conducted on an independent longitudinal dataset of 51 subjects.
• The method successfully synthesized temporal image sequences.

Conclusions:

• The novel deep learning method accurately predicts pediatric development using only cross-sectional training data.
• It enables longitudinal synthesis of clinical images for patient-specific normative development references.
• This approach eliminates the need for longitudinal images in training for personalized pediatric growth prediction.