Bidirectional Projection-Based Multi-Modal Fusion Transformer for Early Detection of Cerebral Palsy in Infants

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Pediatric Neurology

Background:

• Periventricular white matter injury (PWMI) is a common MRI finding in infants with Cerebral Palsy (CP).
• Detecting subtle PWMI lesions in immature infant brains is challenging.
• Early detection of CP is crucial for timely intervention.

Purpose of the Study:

• To develop an AI model for detecting CP in infants under two years old.
• To identify subtle and sparse PWMI lesions using multi-modal MRI data.
• To improve the accuracy and sensitivity of CP diagnosis in infants.

Main Methods:

• Constructed a multi-modal dataset of 243 infant MRI scans (T1WI and T2WI) with annotations for five target regions and lesions.
• Developed a bidirectional projection-based multi-modal fusion transformer (BiP-MFT) model.
• Integrated anatomical T1WI features with T2WI lesion features using a Bidirectional Projection Fusion Module (BPFM).

Main Results:

• The BiP-MFT model achieved a subject-level classification accuracy of 0.90, specificity of 0.87, and sensitivity of 0.94.
• Outperformed nine comparative methods by 0.10 in accuracy, 0.08 in specificity, and 0.09 in sensitivity.
• The BPFM module demonstrated superior performance compared to eight other feature fusion strategies.

Conclusions:

• The developed BiP-MFT model effectively detects Cerebral Palsy in infants by identifying PWMI lesions.
• The multi-modal dataset and fusion transformer approach offer a promising direction for pediatric neuroimaging analysis.
• The study provides a validated method and publicly available resources for advancing CP research.