IEDC‐Net: Internal–External Dual Correction Mechanisms for Semi‐Supervised Medical Image Segmentation

Semi‐supervised learning is crucial in advancing deep learning methods, particularly for addressing the challenges of pixel‐level annotation in medical image segmentation. Limited labeled training data often leads to high uncertainty and inconsistent predictions, especially in complex and ambiguous regions, such as those with adherent tissues or fine branches. These areas lack distinct features, making accurate segmentation extremely difficult. Focusing on such regions can significantly enhance the effectiveness of semi‐supervised training. However, a single model typically suffers from a limited perspective and struggles to detect and correct errors in these challenging areas. To tackle these issues, we propose the internal–external dual correction learning network (IEDC‐Net). IEDC‐Net consists of two independent subnets, each with a shared encoder and two slightly different decoders. Within each subnet, consistency constraints between probabilistic outputs and soft pseudo‐labels are applied, encouraging the model to produce stable, low‐entropy predictions in complex regions. Externally, a difference region inspection module identifies areas of prediction inconsistency, allowing the subnets to mutually supervise and correct errors in these regions. This dual internal–external correction mechanism enables the model to focus on the most challenging areas, ultimately improving performance. We evaluated IEDC‐Net on the left atrium (LA) and pancreas‐CT datasets. Using only 20% of the labeled data for training, our model achieved a Dice score of 90.32% and a Jaccard index of 82.50% on the LA dataset, and a Dice score of 80.40% with a Jaccard index of 67.69% on the pancreas‐CT dataset. These results demonstrate the effectiveness of the proposed method in semi‐supervised medical image segmentation.