Deep learning model stages rectal cancer without manual segmentation on MRI

MST-Net, a segmentation-free deep learning model, achieved AUC 0.95 (accuracy 90%) for T2 vs T3 rectal cancer staging on preoperative T2-weighted MRI, outperforming junior radiologists (69%) and matching senior (85%). External validation AUC 0.85 (accuracy 79%).

• The model uses a pyramid cross-stage feature fusion and a lightweight spatial attention mechanism, eliminating the need for tumor segmentation on T2-weighted MRI.
• Performance was evaluated on an external cohort from a different center, demonstrating generalizability beyond the training data.

Automated summary

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