Locate Then Classify: A Heterogeneous Dual-Teacher Consensus Framework for Semi-Supervised Mammogram Diagnosis

Semi-supervised learning (SSL) offers a pathway to reduce the high cost of pixel-wise annotation in mammography, but classical teacher-student methods suffer from confirmation bias under extreme label scarcity. To address this, we propose a heterogeneous dual-teacher consensus framework (DTSL). Two complementary teachers, one detail-oriented (shallow SE attention) and one semantic-oriented (deep gated modulation), retain pseudo-labels only when they agree, thereby filtering unreliable predictions. Using only 30% of pixel-level labels, DTSL recovers 81.5% of fully supervised segmentation performance (Dice Similarity Coefficient, DSC = 0.626) and achieves a calcification intersection over union (IoU) of 0.700, reducing annotation cost by 70%. It outperforms standard Mean Teacher (78.1%) and nnU-Net plus Mean Teacher (79.8%). The trained segmentation encoder is then transferred to a classification head, injecting morphological priors into malignancy diagnosis. On the INbreast dataset, our classifier attains an AUC of 0.872 and a sensitivity of 0.837, surpassing the average sensitivity of four senior radiologists (0.803) by 3.4 percentage points (p < 0.05). Our model's superior sensitivity suggests its potential as a decision-support tool to reduce false negatives in screening, warranting prospective clinical validation. Grad-CAM visualizations confirm that attention maps align with radiologists' diagnostic cues. The source code is available at https://anonymous.4open.science/r/BrXNetCls-0C14/ .