Search papers, labs, and topics across Lattice.
AI2Department of Pancreatobiliary SurgeryDepartment of RadiologyGuangdong Provincial Clinical ResearchHKUSTImperialNormal UniversityState Key Laboratory of Oncology in SouthSYSUJun 16, 2026arXiv:2606.17972This paper introduces SegDINO, an innovative framework for medical image segmentation that enhances the DINOv3 backbone by integrating lightweight multi-scale structure through Token Pyramid Adaptation (TPA) and Scale-Aware Decoding (SAD). The research highlights that optimizing scale in DINO features is more crucial than merely increasing decoder complexity, leading to significant improvements in efficiency and performance. Extensive evaluations on a newly curated dataset, PanCT, and three public benchmarks reveal that SegDINO achieves state-of-the-art segmentation results while maintaining a lightweight architecture.
Efficiently segmenting small lesions in medical images just got a boost with SegDINO's innovative multi-scale structure, outperforming traditional heavy decoders.
Self-supervised DINO models provide strong transferable visual representations, yet applying them directly to image segmentation remains challenging. Existing approaches commonly rely on heavy decoders with complex upsampling, introducing substantial parameter and computational overhead. We observe that introducing scale into DINO features is far more critical than increasing decoder capacity. In this work, we present SegDINO, an efficient segmentation framework that integrates a DINOv3 backbone with lightweight scale modeling. SegDINO introduces Token Pyramid Adaptation (TPA) to reorganize intermediate DINO features into a pseudo multi-scale hierarchy, and Scale-Aware Decoding (SAD) for efficient intra-scale refinement and top-down multi-scale propagation. We further curate PanCT, a new CT dataset containing 284 patients with expert-annotated pancreatic tumors, to assess SegDINO's ability to handle difficult small-lesion cases. Extensive experiments on PanCT and three public benchmarks demonstrate that SegDINO achieves state-of-the-art results with high efficiency. The code is available at https://github.com/script-Yang/segdino_v2.
