Dynamic layer_view_scale fusion for multimodal medical segmentation

Standard segmentation architectures rely on static fusion to aggregate multi-level features, which often fails to accommodate the heterogeneity of medical images. We aim to design a modality-agnostic decoder that performs input-adaptive fusion across layers, scales, and representational views. We propose DynamicFusionNet, a drop-in decoder framework that learns dense fusion policies conditioned on the input. It uses pixel-wise gating on skip pathways for layer selection, a multi-scale calibration module that integrates complementary branch responses through a shared pointwise projection, and a dual-pathway global-local arbitration design to balance high-level semantics with fine boundary details. These modules can be integrated into standard encoder-decoder pipelines without modifying the backbone. Experiments on five public benchmarks, with main results on BraTS 2021, RETOUCH, and PARSE 2022, show that DynamicFusionNet consistently outperforms strong CNN and Transformer baselines in segmentation accuracy and boundary fidelity across multi-modal and single-modal settings. Combining input-adaptive layer and view fusion with learned multi-scale response calibration provides a unified and effective alternative to static decoder aggregation, improving robustness under heterogeneous imaging conditions. DynamicFusionNet offers a practical, modality-agnostic fusion principle that can be adopted in existing segmentation systems to improve clinical image analysis without bespoke architectures per task. Code: https://github.com/mh-zhou/DynamicFusionNet .