Enhancing medical image segmentation with HCE-Net: A heterogeneous feature collaborative extraction network

Medical image segmentation occupies a pivotal role in clinical diagnosis and therapeutic planning. Conventional segmentation techniques struggle to accommodate the intricate representational attributes of medical images, while prevailing deep learning segmentation models predominantly adopt homogeneous single encoders and linear feature fusion paradigms, thereby exhibiting inherent limitations in contextual modeling, fine-grained feature excavation, and contour segmentation accuracy. To address such dilemmas, this paper proposes a Heterogeneous Collaborative Feature Extraction Network (HCE-Net). Establishing a dual-path heterogeneous encoding architecture, this network excavates complementary feature representations in a parallel manner. An Orthogonal Fusion Module (OFM) is elaborately devised to integrate Sobel edge priors with multi-scale convolutions, enabling the precise capture of edge and contour characteristics. A Diversified Receptive Field Module (DFM) is constructed to consolidate multi-scale detail manifestations and preserve global contextual cues. Furthermore, Adaptive Aggregation Connection (AAC) is embedded within skip connection branches to supplant conventional simplistic concatenation, efficaciously mitigating the technical bottleneck of heterogeneous feature fusion. Comparative experiments conducted on three publicly accessible medical datasets, encompassing dental radiographs, pulmonary CT scans, and retinal vascular images, demonstrate that HCE-Net substantially elevates segmentation accuracy, enhances boundary integrity, and ameliorates the general robustness of the model. It manifests distinguished application prospects across diverse medical image segmentation scenarios.