Research map / Detection and segmentation / Open-set and few-shot detection
RAM-W600: A Multi-Task Wrist Dataset and Benchmark for Rheumatoid Arthritis
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.
Introduces RAM-W600, a public multi-task wrist conventional radiography dataset with instance-level wrist bone segmentation masks and SvdH bone erosion scores, plus benchmarks of segmentation and BE classification models.
Problem
Computer-aided diagnosis of rheumatoid arthritis at the wrist is limited by the lack of public, high-quality instance-level bone annotations and SvdH bone erosion scores in conventional radiographs.
Approach
A wrist-focused conventional radiography dataset that pairs pixel-level instance segmentation of 14 wrist bones (618 images) with SvdH BE scores for six joint surfaces (800 RA images, 4800 ROIs), and benchmarks supervised, foundation, and classification models on it.
Key results
- SwinUMamba achieved the best segmentation performance with 97.75% DSC and 90.71% NSD on the test set, while SAM and MedSAM reached lower DSC (up to 88.7% and 85.1%).
- For BE vs nonBE classification, the best balanced accuracy was 52.64% (MobileViT) and best F1 was 12.40% (EfficientFormer), indicating the task is hard.
- Models with high specificity (ConvKAN 94.70%) had very low sensitivity (3.82%), reflecting strong bias toward negative predictions under class imbalance.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Evaluate segmentation on BE-affected rather than nonBE radiographs | Most supervised segmentation models · RAM-W600 test set | DSC | Hurt |
| Use general-purpose foundation segmentation models instead of supervised models | Wrist bone segmentation · RAM-W600 test set | DSC | Hurt |
Surprising findings
- Foundation models SAM and MedSAM underperform supervised models on wrist bone segmentation despite their general-purpose pretraining.
- Segmentation DSC differs significantly between BE and nonBE samples for most models, but NSD does not.
Limitations
- RA cases come from a single geographic region and a relatively homogeneous ethnic population, limiting generalizability.
- SvdH BE score distribution is highly imbalanced, with high scores underrepresented.
- Advanced RA cases with bony ankylosis (430 images) were excluded from the segmentation subset.
Experimental setup
Datasets: RAM-W600 Digital Hand Atlas BTXRD FracAtlas
Benchmarks: RAM-W600
Metrics: DSC NSD VOE MSD RAVD BACC F1-score DOR ACC SEN
Baselines: UNet DeepLabV3 FPN PSPNet DeepLabV3+ SegResNet UNet++ SegFormer TransUNet UKAN UMambaBot UMambaEnc
Builds on: SvdH scoring system Mamba-based U-Net variants
Closest papers by meaning
- MosaicMRI: A Diverse Dataset and Benchmark for Raw Musculoskeletal MRI (NeurIPS 2026 · similarity 0.83)
- VidSole: A Multimodal Dataset for Joint Kinetics Quantification and Disease Detection with Deep Learning (AAAI 2025 · similarity 0.81)
- BLS-GAN: A Deep Layer Separation Framework for Eliminating Bone Overlap in Conventional Radiographs (AAAI 2025 · similarity 0.80)
- PolyPose: Deformable 2D/3D Registration via Polyrigid Transformations (NeurIPS 2025 · similarity 0.78)
- MedicalNarratives: Connecting Medical Vision and Language with Localized Narratives (NeurIPS 2025 · similarity 0.77)
- HIT: Estimating Internal Human Implicit Tissues from the Body Surface (CVPR 2024 · similarity 0.77)
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.