Research map / Detection and segmentation / Open-set and few-shot detection

RADAR: Text-Guided Medical Image Segmentation via Residual Aggregation and Dense Alignment Representations

NeurIPS 2026 · Detection and segmentation › Open-set and few-shot detection · approach: Open-vocabulary semantic segmentation

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes RADAR, a parameter-efficient text-guided medical segmentation framework that couples a frozen BiomedCLIP (LoRA-adapted) with MedSAM2 via text-conditioned depth aggregation, structured dense alignment, and composite prompt generation, reaching strong results with 0.74M trainable parameters.

Problem

Text-guided medical image segmentation struggles to convert high-level textual semantics into accurate pixel-level spatial prompts, especially for low-contrast lesions, ambiguous boundaries, and similar anatomical contexts.

Approach

Uses text as a depth-routing condition to aggregate historical cross-layer visual features (GRDA), aligns crop-level target/boundary/context dense features of BiomedCLIP to a frozen teacher with a boundary negative rejection term (SDRA), and converts the resulting similarity map into text, point, and mask prompts for MedSAM2 (MPG).

Key results

• RADAR reaches 84.12% Dice on QaTa-COV19, 75.47% on MosMedData+, and 92.15% on Kvasir with 0.74M trainable parameters.
• Improves Dice by 6.65, 10.12, and 5.96 points over the strongest pure visual baselines on QaTa-COV19, MosMedData+, and Kvasir respectively.
• Removing GRDA or SDRA degrades Dice on MosMedData+ (74.63 and 73.71 vs 75.47 for full RADAR in ablation).

Surprising findings

• Learnable prompt generation (RADAR-LMPG) slightly improves Kvasir Dice and mIoU but is worse on MosMedData+, so handcrafted MPG offers a better parameter-performance trade-off in this setting.

Limitations

• Validation limited to 2D tasks; no native 3D, multi-organ, or multiclass evaluation.
• MPG relies on heuristic thresholding, connected-component analysis, and morphological operations that may be unstable under low contrast or inaccurate text.
• Performance drops notably with empty or incorrect text, showing dependence on correct image-text correspondence.

Experimental setup

Datasets: QaTa-COV19 MosMedData+ Kvasir

Benchmarks: QaTa-COV19 MosMedData+ Kvasir

Metrics: Dice mIoU HD95 ASSD

Baselines: U-Net UNet++ Swin-UNet TransUNet LGMS LViT MMI-UNet LAVT RecLMIS MTGT HiMix DIT

Builds on: LoRA MedSAM2 BiomedCLIP Attention Residuals CLIPSelf

Closest papers by meaning

• S2S2: Semantic Stacking for Robust Semantic Segmentation in Medical Imaging (AAAI 2025 · similarity 0.86)
• Unleashing the Potential of SAM for Medical Adaptation via Hierarchical Decoding (CVPR 2024 · similarity 0.85)
• CAT: Coordinating Anatomical-Textual Prompts for Multi-Organ and Tumor Segmentation (NeurIPS 2024 · similarity 0.85)
• SGTC: Semantic-Guided Triplet Co-training for Sparsely Annotated Semi-Supervised Medical Image Segmentation (AAAI 2025 · similarity 0.84)
• BiomedCoOp: Learning to Prompt for Biomedical Vision-Language Models (CVPR 2025 · similarity 0.84)
• Enhancing SAM with Efficient Prompting and Preference Optimization for Semi-supervised Medical Image Segmentation (CVPR 2025 · similarity 0.84)