GloResNet: A lightweight 3D CNN with global topological features for preterm brain injury prediction

Researchers introduced GloResNet, a lightweight 3D CNN for predicting brain injury in preterm infants from T2-weighted MRI.

The model combines a ResNet-10 backbone pretrained on MedicalNet with global manifold mapping, z-score normalization, mixup, class weighting, and test-time augmentation.

On the dHCP dataset, GloResNet achieved 75.18% average accuracy in 5-fold cross-validation, with 0.81 specificity and 0.76 sensitivity.

The work points to a non-invasive MRI-based screening tool that could help identify at-risk infants earlier and support timely intervention.