Dual-teacher semi-supervised model cuts mammogram labeling need by 70% and tops radiologists' sensitivity for calcifications

A dual-teacher semi-supervised model using only 30% pixel labels recovered 81.5% of fully supervised segmentation (Dice 0.626) and calcification IoU 0.700. On INbreast, its classifier had AUC 0.872 and sensitivity 0.837 vs senior radiologists' 0.803 (p<0.05).

• Using 30% of pixel-level labels, the segmentation model achieved Dice 0.626, calcification IoU 0.700, and 81.5% of fully supervised performance, outperforming Mean Teacher (78.1%) and nnU-Net plus Mean Teacher (79.8%).
• Transferred to malignancy classification, it reached AUC 0.872 and sensitivity 0.837 on INbreast, exceeding the average senior radiologist sensitivity of 0.803 by 3.4 percentage points (p<0.05).
• Grad-CAM attention maps aligned with radiologists' diagnostic cues; the authors call for prospective clinical validation.

Automated summary

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