3D Mamba-CNN Framework Boosts Multi-Center NPC MRI Segmentation

MambaGLA-NPC, a 3D model combining CNNs and state space models, segments nasopharyngeal carcinoma on MRI with Dice 80.47% internally and 87.13%/76.70% on two external sets, outperforming prior methods.

• On internal test, Dice 80.47%, IoU 67.83%, HD95 4.09 mm; cross-center Dice gains up to 8.84% with NPC-Adapter.
• Conv-GLSSM module reduces computational cost vs Transformers (21.66 GFLOPs vs out-of-memory for self-attention at 128³ resolution).

Automated summary

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