Accurate medical image segmentation requires simultaneous modeling of robust global semantics and precise anatomical boundaries. While Vision Transformers (ViTs) exhibit strong global representation capability, they often struggle to capture fine-grained local structures and boundary details, particularly in scenarios with low contrast and blurred boundaries. To address this limitation, we propose Glasses, a novel framework that equips pre-trained Transformers with a structure-enhanced visual adaptation mechanism for medical image segmentation. Specifically, Glasses employs a frozen DINOv3 as the backbone encoder, with only the terminal layers fine-tuned. We introduce a Multi-Scale Deformable Attention Adapter that performs cross-scale interaction between lightweight convolutional structural priors and Transformer tokens extracted from specific intermediate layers of DINOv3, thereby transforming single-scale token embeddings into hierarchical representations enriched with both local structural cues and global context. This design preserves the stability of pre-trained semantic representations while improving the modeling of local anatomical structures and scale variations. Furthermore, we introduce an independent Edge-Aware Convolutional Encoder that explicitly learns boundary-enriched multi-scale representations under nested boundary supervision, providing a strong source of fine-grained local structural information that complements the adapter-enhanced Transformer features during decoding. A U-Net-style decoder then progressively fuses the adapter-enhanced Transformer features and the boundary-enriched convolutional representations to recover high-resolution segmentation outputs. Extensive experiments on the Synapse and ISIC datasets demonstrate that Glasses achieves Dice coefficients of 83.52 https://github.com/friay/Glasses .
Vessel segmentation in digital subtraction angiography (DSA) is crucial for intraoperative visualization and quantitative assessment during endovascular interventions. However, accurate segmentation remains challenging due to complex thin vascular branches and strong background interference from residual anatomical structures after subtraction. Although recent methods incorporate temporal modeling using CNNs, recurrent networks, or transformers, they often treat temporal information implicitly and fail to explicitly exploit the contrast flow dynamics in DSA sequences. In this work, we present a spatiotemporal DSA vessel segmentation framework that combines a forward phase-difference prior with Mamba-based temporal sequence modeling. The Mamba-based temporal encoder models long-range temporal dependencies and contrast flow dynamics with linear computational complexity. To capture contrast-driven intensity changes, we introduce a forward phase-difference projection that enhances contrast-responsive vascular structures while suppressing temporally stable background tissues. A minimum intensity projection (MinIP) is further used as a complementary static spatial prior to provide coarse vessel structural cues. These priors are integrated with temporal features through a lightweight domain-aware fusion module before decoding. Extensive experiments on the DIAS and DSCA datasets demonstrate that the proposed method consistently outperforms state-of-the-art approaches, particularly in preserving vascular connectivity and segmenting thin distal vessels. Code is available at https://github.com/kkk123lll/FPD-SegDSA .
In recent medical image analysis, Convolutional Neural Networks (CNNs) and State Space Models (SSMs) have set benchmarks in segmentation tasks. While CNNs excel in capturing local fine-grained features, SSMs achieve remarkable global context understanding with linear complexity. However, Mamba-UNet, a pioneering pure SSM-based model, still exhibits deficiencies in feature representation, fusion efficiency, and spatial detail reconstruction. To address these limitations, we propose VM-NeXT UNet, an improved architecture that synergizes ConvNeXT with VSS Blocks. VM-NeXT UNet introduces four key optimizations: (1) a dual-encoder parallel structure for comprehensive feature extraction; (2) a channel-spatial attention gating module in skip connections for adaptive feature screening; (3) multi-scale convolution layers in the decoder to preserve spatial details; and (4) a combined FocalLoss and DiceLoss strategy to focus on hard samples. Experiments on the Synapse and ACDC datasets yielded Dice scores of 86.21
Due to structural anisotropy and complex anatomical boundaries, 3D medical image segmentation remains a challenging task. To address these issues, this paper proposes the Direction-Sensitive Equilibrium Mamba Network (DEM-Mamba). Through directional decoupling, balanced enhancement, and adaptive fusion, this network systematically enhances structure perception and boundary preservation capabilities in 3D space. Specifically, DEM-Mamba integrates three core designs: First, it employs an Orientation-Aware Module (OAM) based on learnable Riesz wavelets to decouple high-frequency directional features and mitigate information imbalance caused by anisotropic resolution. Second, it designs a Bidirectional Balanced Mamba (BBM), which effectively captures long-range anisotropic dependencies and maintains boundary clarity through complementary bidirectional scanning combined with adaptive normalization. Finally, it introduces an Orientation-Conflict Fusion (OCF) mechanism that dynamically aggregates multi-directional information based on correlation weights, enhancing the model’s adaptability to complex boundaries. Experiments across multiple 3D medical image datasets demonstrate that this method outperforms existing state-of-the-art approaches in both segmentation accuracy and boundary integrity, providing an effective new framework for 3D medical image segmentation. Code is available at https://github.com/pky888/DEM-Mamba
