Three-phase learning framework for first-trimester fetal echocardiography plan classification via knowledge distillation and CBAM-enhanced DenseNet

Early detection of fetal heart defects can support timely clinical interventions, thereby reducing neonatal mortality. Deep learning (DL) has been widely applied to fetal echocardiography image classification (FEIC) to enhance the diagnosis and detection of heart defects. However, at the early pregnancy stage (12–14 weeks), underdeveloped fetal cardiac structures and even slight probe movement can also cause significant variations in view appearance, as a result leading to challenges for DL models in accurately recognizing complex anatomical structures. We propose a novel three-phase progressive learning framework built upon DenseNet-201 backbone, enhanced with a convolutional block attention module (CBAM) and knowledge distillation (KD) for effectively addressing the early-stage FEIC problem. The key novelty lies in a staged learning strategy that incrementally improves model generalization from simple to complex visual patterns, rather than relying solely on architectural modifications. Firstly, we train a convolutional neural network (CNN) to separate the minority X-sign class into clear and deformed subsets. In Phase 1, DenseNet-201 is fine-tuned on clear samples to learn structured features. In Phase 2, the model is trained on deformed samples with KD to transfer prior knowledge and improve robustness. Finally, Phase 3 trains the model on the full dataset to enhance generalization. Experimental results show consistent improvements over state-of-the-art (SOTA) baselines, with gains of 3% and 1% overall, and notable improvements of about 5% (recall) and 3% (F1) for the X-sign class. These results demonstrate the effectiveness of the progressive learning strategy in handling class imbalance and structural variability.