Recurrent group-equivariant nnU-Net boosts brain tissue segmentation and early stroke detection on non-contrast CT

RGCNN-nnUNet achieved Dice scores of 0.885 (white matter) and 0.86 (deep gray matter) on multi-center NCCT. When its tissue-specific Z-score maps were combined with AI-denoised CT, a blinded reader study showed a 29% increase in diagnostic sensitivity for stroke compared with st…

• The architecture introduces a Recurrent Group Convolutional Cell that leverages group equivariance to provide consistent anatomical features and stable iterative refinement during segmentation.
• Improvements over baseline included a 14.3% and 10.6% reduction in Hausdorff Distance (HD95) for the segmented structures.
• This is a technical development study with a single-reader comparison; generalizability to broader clinical practice and multiple readers remains to be established.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.