Structural entropy driven dynamic fusion for incomplete multi-view myocardial disease detection

Early and accurate detection of myocardial infarction is crucial for improving patient outcomes. Multi-view echocardiography provides complementary information on cardiac structure and function from different planes. However, in real-world clinical settings, issues such as missing views and inconsistent view reliability are common. Existing incomplete multi-view learning methods often rely on missing view completion or implicit fusion mechanisms, which can easily introduce reconstruction bias and additional computational overhead, and make it difficult to explicitly capture the quality differences between visible views. To this end, this paper proposes a Structural Entropy-driven Dynamic Fusion (SEDF) method, which performs feature learning and diagnostic decision-making using only the visible views. This method evaluates the stability of a view’s internal structure and the reliability of its information using structural entropy, and maps these metrics to dynamic fusion weights at the sample level. This approach adaptively enhances the contribution of highly reliable views while suppressing interference from low-quality views. By further integrating cross-view consistency constraints, structure-entropy-guided contrastive learning, and weight-entropy regularization, we enhance the stability, discriminative power, and generalization ability of the fused representation. Experimental results on the HMC-QU myocardial infarction dataset and two public datasets, ADNI and ROSMAP, show that the proposed method outperforms representative incomplete multi-view learning methods and demonstrates good robustness and interpretability under conditions of missing views and heterogeneous image quality.