CC-Net: A cross-hierarchical context-aware network for medical image segmentation

Accurate segmentation of lesion regions in medical images is crucial for early diagnosis and precise treatment of diseases. Existing U-shaped segmentation networks typically employ direct connections between features of the same hierarchical level in the encoder-decoder structure; however, the lack of explicit interaction between different hierarchical features limits the effective integration of shallow structural information and deep semantic features, resulting in challenges when delineating complex boundaries and fine-grained structures. To address these challenges, we propose a Cross-Hierarchical Context-Aware Network (CC-Net). The model comprises four key components: The Cross-Hierarchical Feature Aggregation (CFA) Module aggregates shallow and deep features separately, explicitly modeling semantic relationships across hierarchical levels. The Global Feature Aggregation (GFA) Module integrates features from multiple levels to construct a unified global representation, providing multi-level semantic context during cross-branch feature interactions. The Cross-Branch Semantic Supplement (CSS) Module injects the global semantic information obtained from GFA into the CFA-aggregated shallow and deep features, enhancing overall structural awareness. The Enhanced Feature (EF) Module strengthens the model's discriminative capability for target regions while effectively suppressing background noise. Extensive experiments on six public medical image datasets demonstrate that CC-Net consistently outperforms state-of-the-art segmentation methods across all evaluation metrics. The code is available at https://github.com/zz0226zz/CC-Net.