Research map / Detection and segmentation / Open-set and few-shot detection
iDPA: Instance Decoupled Prompt Attention for Incremental Medical Object Detection
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 iDPA, a prompt-based continual learning framework for incremental medical object detection that generates instance-level prompts and decouples prompt attention in GLIP, evaluated on a new 13-dataset medical benchmark ODinM-13.
Problem
Continual learning for medical object detection with vision-language detectors: prompt-based CL methods designed for classification couple foreground/background information and dilute prompts against long image-text token sequences, causing poor localization and catastrophic forgetting across hetero
Approach
Instance-Level Prompt Generation (IPG) builds prompts from RoI-pooled instance features via cross-attention, and Decoupled Prompt Attention (DPA) separates prompt-to-vision and prompt-to-text knowledge injection from the original attention coupling, applied at the multimodal fusion encoder of GLIP.
Key results
- On ODinM-13 full-data, iDPA reaches 50.28 FAP and 2.48 FFP, improving FAP by 5.44 over the best prompt-based baseline NoRGa.
- In the 10-shot setting iDPA achieves 23.78 FAP, 12.88 points above the best alternative.
- iDPA uses about 1.4% of trainable parameters (3.22M) relative to full fine-tuning and reports lower memory than the prompt baselines.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use iDPA instead of NoRGa | GLIP with Swin-T · ODinM-13 full data | FAP: 5.44% | Helped |
| use iDPA instead of DualPrompt | GLIP with Swin-T · ODinM-13 1-shot | FAP: 4.38% | Helped |
| use iDPA instead of S-Prompt | GLIP with Swin-T · ODinM-13 10-shot | FAP: 12.88% | Helped |
| use iDPA instead of DIKI | GLIP with Swin-T · ODinM-13 50-shot | FAP: 4.59% | Helped |
| add IPG with weight transfer to naïve prompts (ablation) | iDPA · ODinM-13 full data | FAP: 3.66% | Helped |
| add DPA to naïve prompts (ablation) | iDPA · ODinM-13 full data | FAP: 2.11% | Helped |
| combine IPG with weight transfer and DPA (ablation) | iDPA · ODinM-13 full data | FAP: 5.29% | Helped |
| remove weight transfer from IPG (ablation) | iDPA · ODinM-13 full data | FAP | Hurt |
| inject prompts in fusion encoder rather than vision encoder only (ablation) | iDPA · ODinM-13 full data | FAP | Helped |
| inject prompts in all three encoders rather than fusion encoder only (ablation) | iDPA · ODinM-13 full data | FAP | Helped |
| use 6 fusion attention layers instead of fewer layers (ablation) | iDPA · ODinM-13 full data | FAP | Helped |
| set RoI scale γ to 1.30 rather than 1.00 or 1.50 (ablation) | iDPA · ODinM-13 full data | FAP | Helped |
Surprising findings
- Naive sequential fine-tuning and ZiRa almost completely forget earlier tasks on several datasets (FAP 4.40 and 3.66), while the ER baseline with replay performs much better.
Limitations
- Evaluated only on GLIP with Swin-T backbone; generality to other VLOD models is not shown.
- Few-shot and polyp results come from a single backbone and a limited set of task orders (three seeds).
Experimental setup
Datasets: DFUC-2020 Kvasir OpticNerv BCCD CPM-17 BreastCancer TBX11K KidneyTumor Luna16 ADNI Meningioma BreastTumor
Benchmarks: ODinM-13 four polyp datasets (domain continual setting)
Metrics: FAP CAP FFP AP AP50 AP75 parameters memory training time
Baselines: L2P DualPrompt CODA S-Prompt DIKI NoRGa ER WiSE-FT ZiRa Sequential SAM2.1 MedSAM-2
Builds on: GLIP L2P DualPrompt VPT-Deep CODA-Prompt
Closest papers by meaning
- Prompt as Knowledge Bank: Boost Vision-language model via Structural Representation for zero-shot medical detection (ICLR 2025 · similarity 0.85)
- BiomedCoOp: Learning to Prompt for Biomedical Vision-Language Models (CVPR 2025 · similarity 0.85)
- One-Prompt to Segment All Medical Images (CVPR 2024 · similarity 0.84)
- GRAPE: Graph-Augmented Prototype Explanations for Interactive Medical Image Diagnosis (NeurIPS 2026 · similarity 0.84)
- Memory-Efficient Prompt Tuning for Incremental Histopathology Classification (AAAI 2024 · similarity 0.84)
- Efficient Deformable Convolutional Prompt for Continual Test-Time Adaptation in Medical Image Segmentation (AAAI 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.