DTEF ‐Net: A Dual‐Teacher and Entropy‐Guided Semi‐Supervised Framework With Multiperspective Copy‐Paste for Medical Image Segmentation

Accurate medical image segmentation is crucial for computer‐aided diagnosis; however, the scarcity of annotated data presents a significant challenge. Traditional student–teacher frameworks address this by generating pseudo‐labels from the teacher model. Nevertheless, these pseudo‐labels are often plagued by high noise and lack of diversity. To tackle this issue, we propose a novel semi‐supervised segmentation framework, DTEF‐Net, which integrates the Dual‐Teacher Prediction Weighted Fusion Module (PEM), the Multiperspective Fusion Bidirectional Copy‐Paste strategy (MFBCP), and the Fusion Entropy‐Modulated Consistency Loss (FECL). PEM fuses the predictions of two independent teacher models using an entropy‐based adaptive weighting approach, thereby enhancing the reliability and diversity of pseudo‐labels. MFBCP leverages predictions from multiple teachers to improve the accuracy of foreground‐background exchanges, providing stronger supervisory signals for the student model. FECL dynamically adjusts consistency constraints at the voxel level using entropy, particularly enhancing robustness and segmentation precision in uncertain regions. Extensive experiments on the LA and Pancreas‐CT datasets demonstrate that DTEF‐Net significantly outperforms state‐of‐the‐art methods in scenarios with limited labeled data, offering a promising direction for efficiently utilizing unlabeled data in medical image segmentation.