Automating early prediction of cerebral palsy: A transfer learning model for infant MRI analysis

Motivation: Clinical diagnosis of cerebral palsy (CP) is often delayed, costly, and limited in accuracy, highlighting the need for a low-cost, automated, early diagnosis approach. Goal(s): This study develops a deep transfer learning (TL) model for early CP prediction in infants aged 6 months to 2 years. Approach: The CP-TL model, based on ResNet-18 with two pre-trained weights, was tested on various dataset combinations. Performance was evaluated using accuracy, sensitivity, specificity, and AUC. Results: The single-center dataset with pre-trained weight 1 achieved the best performance (accuracy: 92.19%, AUC: 0.9655), while weight 2 excelled with multicenter data, highlighting the need for weight selection. Impact: This model demonstrates the potential of deep transfer learning for early CP prediction, offering reliable support for early intervention and rehabilitation planning in infants aged 6 months to 2 years, with significant clinical application value.