Co-learning-assisted progressive dense fusion network for cardiovascular disease detection using ECG and PCG signals

Abstract

Electrocardiograms (ECGs) and phonocardiograms (PCGs) are two modalities to provide complementary diagnostic information for improving the early detection accuracy of cardiovascular diseases (CVDs). Existing multi-modality methods mainly used the early or late feature fusion strategy, which did not simultaneously utilize the complementary information contained in low-level detail features and high-level semantic features of different modalities. Meanwhile, they were specially designed for the multi-modality scenario with both ECGs and PCGs, without considering the missing-modality scenarios with only ECGs or PCGs in clinical practice. To address these challenges, we developed a Co-learning-assisted Progressive Dense fusion network (CPDNet) for end-to-end CVD detection, with a three-branch interweaving architecture consisting of ECG and PCG modality-specific encoders and a progressive dense fusion encoder, which could be used for both multi-modality and missing-modality scenarios. Specifically, we designed a novel progressive dense fusion strategy, which not only progressively fused multi-level complementary information of different modalities from low-level details to high-level semantics, but also employed the dense fusion during feature fusion at each level to further enrich available multi-modality information through mutual guidance of features at different levels. Meanwhile, the strategy integrated cross-modality region-aware and multi-scale feature optimization modules to fully evaluate the contributions of different modalities and signal regions and enhance the feature extraction ability of the network for multi-scale target regions. Moreover, we designed a novel co-learning strategy to guide the learning process of the CPDNet by combining intra-modality and joint losses, which made each encoder well-trained. This strategy could not only assist our fusion strategy by making modality-specific encoders provide sufficiently discriminative features for the fusion encoder, but also enable the CPDNet to robustly handle missing-modality scenarios by independently using the corresponding modality-specific encoder. Experimental results on public and private datasets demonstrated that our method not only outperformed state-of-the-art multi-modality methods by at least 5.05% for average accuracy in the multi-modality scenario, but also achieved better performance than single-modality models in the missing-modality scenarios.