Title: GloResNet: A Lightweight 3D CNN Integrating Global Topological Features for Predicting Preterm Brain Injury

Abstract:

This research presents an automated deep learning framework designed to predict brain injury (BI) in premature infants using T2-weighted MRI scans from the dHCP dataset. To tackle the challenge of limited data availability, we introduce GloResNet, a resource-efficient 3D CNN built upon the ResNet-10 architecture and initialized with weights from MedicalNet. The methodology employs a global manifold mapping approach, which first resamples every 3D volume to dimensions of 128x128x128 and subsequently applies subject-wise z-score intensity normalization. This process ensures that global topological structures are maintained while standardizing image appearance. To enhance model robustness, the training regimen incorporates mixup techniques, class weighting, and test-time augmentation. Evaluation via 5-fold cross-validation yielded an average accuracy of 75.18%, reaching a peak of 81.82%, alongside specificity and sensitivity scores of 0.81 and 0.76, respectively. These findings confirm that a lightweight, topology-sensitive CNN can effectively predict neonatal brain injury, serving as a viable non-invasive screening solution. The code for this study is publicly available at: https://github.com/ICL-SUST/GloResNet-Preterm-Brain

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC