Atlas is Your Perfect Context: One-Shot Customization for Generalizable Foundational Medical Image Segmentation

TL;DR

AtlasSegFM tackles the challenge of generalizable medical image segmentation across diverse clinical contexts by enabling one-shot customization of segmentation foundation models. It fuses a context-rich atlas-derived prompt via test-time registration with powerful foundation models, and refines results through a learnable fusion mechanism. The approach achieves state-of-the-art or competitive results across six datasets, including small and delicate structures, while remaining efficient for clinical workflows. Its lightweight, training-free adaptation and robust out-of-distribution performance underscore its practical impact for real-world radiology and therapy planning.

Abstract

Accurate medical image segmentation is essential for clinical diagnosis and treatment planning. While recent interactive foundation models (e.g., nnInteractive) enhance generalization through large-scale multimodal pretraining, they still depend on precise prompts and often perform below expectations in contexts that are underrepresented in their training data. We present AtlasSegFM, an atlas-guided framework that customizes available foundation models to clinical contexts with a single annotated example. The core innovations are: 1) a pipeline that provides context-aware prompts for foundation models via registration between a context atlas and query images, and 2) a test-time adapter to fuse predictions from both atlas registration and the foundation model. Extensive experiments across public and in-house datasets spanning multiple modalities and organs demonstrate that AtlasSegFM consistently improves segmentation, particularly for small, delicate structures. AtlasSegFM provides a lightweight, deployable solution one-shot customization of foundation models in real-world clinical workflows. The code will be made publicly available.

Figures (9)

• Figure 1: Comparison of three segmentation paradigms. Our method adapts off-the-shelf segmentation foundation models to each target context using one atlas. "Uncommon-scene Dice" averages BrainRT (organs-at-risk) and Fe-MRA (vessels).
• Figure 2: Visual comparisons with recent interactive method (i.e., nnInteractive isensee2025_arxiv_nninteractive) on the BrainRT dataset for radiotherapy. nnInteractive performs far below expectation in context #1 (organs-at-risk), an uncommon contexts underrepresented in its training data.
• Figure 3: Pipeline including three steps: 1) Registration between query and support to obtain the mask (see Sec. \ref{['sec:registion']}), 2) Prompting the foundation model based on the mask from atlas (see Sec. \ref{['sec:foundation_model']}), and 3) Fusion of the two masks to obtain the final result (see Sec. \ref{['sec:fusion']}). Our model uses an inference-only design, where the "fire" denotes test-time adaptation and the "snow" remains frozen.
• Figure 4: Visual comparisons against recent one-shot learning methods with open-source implementations (i.e., UniverSeg butoi2023_ICCV_universeg, Tyche rakic2024cvpr_tyche, and Iris gao2025_CVPR_Iris) across four datasets. Each color corresponds to a specific context, and all contexts are combined in the visualization.
• Figure 5: Qualitative results of AtlasSegFM with nnInteractive using atlas-derived prompts on Abd-CT. Rows indicate prompt type (Click / Box / Mask); columns show Spleen, Right Kidney, Left Kidney, and Liver. Dice is reported per case. The Mask prompt is the most robust, particularly for challenging targets such as kidneys by providing a dense structural prior beyond coarse localization.