RAM-W600: A Multi-Task Wrist Dataset and Benchmark for Rheumatoid Arthritis¶
Conference: NeurIPS 2025 arXiv: 2507.05193 Code: GitHub Area: Medical Imaging Keywords: Rheumatoid Arthritis, Carpal Bone Segmentation, Bone Erosion Scoring, Dataset, Instance Segmentation
TL;DR¶
RAM-W600 is the first publicly available multi-task wrist conventional radiograph dataset, comprising 1,048 images and supporting two clinically relevant tasks: carpal bone instance segmentation and SvdH bone erosion (BE) scoring, accompanied by comprehensive benchmarking.
Background & Motivation¶
Rheumatoid arthritis (RA) is a prevalent autoimmune disease in which the wrist joint serves as a central region for diagnosis. Conventional radiography (CR) is widely used in clinical practice for disease screening and assessment due to its low cost and accessibility. However, computer-aided diagnosis (CAD) research for carpal bones remains severely constrained, with the primary bottleneck being the extreme difficulty of obtaining high-quality instance-level annotations:
Anatomical complexity: The wrist comprises multiple small bones with narrow joint spaces, complex overlapping structures, and frequent occlusions, requiring substantial anatomical expertise for accurate annotation.
Pathological interference: RA progression introduces osteophytes, bone erosions (BE), and even bony ankylosis, which alter bone morphology and further increase annotation difficulty.
Limitations of existing datasets: Publicly available hand radiograph datasets either lack pixel-level segmentation annotations for the wrist or provide incomplete BE scores, rendering them insufficient for RA research.
Existing work has predominantly focused on carpal bone segmentation in CT/MRI modalities, with small-scale datasets (mostly private, ranging from tens to hundreds of images). CR-based carpal bone segmentation research is scarce, particularly lacking publicly available datasets suited to complex pathological conditions.
The paper's Key Insight is to construct a multi-task, multi-center carpal bone CR benchmark dataset that simultaneously covers two clinically critical tasks—instance segmentation and SvdH BE scoring—thereby lowering the barrier to entry for RA wrist research.
Method¶
Overall Architecture¶
RAM-W600 does not propose a new model; rather, it contributes a dataset and benchmark. The dataset design encompasses a complete pipeline: data collection → image preprocessing → expert annotation → data splitting → benchmark evaluation.
Key Designs¶
- Dataset Composition and Annotation Framework
The dataset includes 1,048 wrist CR images from 388 patients across 6 medical centers. Among these, 618 images provide pixel-level instance segmentation annotations covering 14 carpal bone categories, and 800 images include SvdH BE scores across 6 joint surfaces (4,800 scores in total).
The annotation framework consists of three levels: - Anatomical structure annotation: Precise contour delineation of 14 carpal bones (MC1–5, Tr, Tz, Sca, Lu, Cap, Ham, Tri, Radius, Ulna), with each bone independently annotated using a multi-label strategy. - Bone location annotation: ROI annotation for 6 joint regions of interest defined by the SvdH scoring system. - SvdH BE scoring annotation: BE severity scoring for 6 key joint surfaces.
- Data Diversity and Quality Control
Data sources include 3 Japanese medical institutions (HMCRD, SCGH, HU) and 3 public datasets (DHA, BTXRD, FA), covering 207 RA patients and 181 non-RA patients. Image parameters are managed via the DICOM standard at a resolution of 0.15 mm/pixel (internal cohort).
The annotation pipeline was supervised by senior radiologists and employed rigorous review procedures to ensure quality. The dataset design emphasizes stratified splitting of BE and non-BE cases to ensure representativeness across training and test sets.
-
Benchmark Evaluation Protocol
-
Segmentation task: Evaluation of 13 supervised models (UNet, DeepLabV3+, TransUNet, SwinUMamba, etc.) and 3 foundation models (SAM, MedSAM), using DSC, NSD, VOE, MSD, and RAVD metrics.
- BE classification task: Evaluation of 7 classification models (MobileViT, ResNet, MedMamba, etc.), using BACC, F1, DOR, ACC, SEN, SPC, and PRE metrics.
Loss & Training¶
All benchmark experiments adopt the AdamW optimizer (weight decay = 1e-2) with cosine annealing learning rate scheduling. For the segmentation task: initial learning rate 1e-4, 100 epochs, batch size 8; for the classification task: initial learning rate 1e-6, 100 epochs, batch size 16. All experiments were repeated 5 times with fixed seeds on an RTX 4090.
Key Experimental Results¶
Main Results — Segmentation¶
| Model | DSC (%) ↑ | NSD (%) ↑ | VOE (%) ↓ | MSD (pix) ↓ | Params |
|---|---|---|---|---|---|
| SwinUMamba | 97.75 | 90.71 | 4.35 | 1.06 | 59.89M |
| TransUNet | 97.62 | 89.48 | 4.60 | 1.05 | 105.91M |
| UMambaEnc | 97.56 | 89.10 | 4.71 | 1.11 | 4.58M |
| Unet++ | 97.33 | 86.99 | 5.15 | 1.36 | 2.41M |
| SAM (box) | 88.74 | 64.40 | 18.45 | 4.25 | 641.09M |
| MedSAM (box) | 85.07 | 38.81 | 25.15 | 5.97 | 93.74M |
Main Results — BE Classification¶
| Model | BACC (%) | F1 (%) | DOR | SEN (%) | SPC (%) |
|---|---|---|---|---|---|
| MobileViT | 52.64 | 11.85 | 1.82 | 21.06 | 84.23 |
| EfficientFormer | 50.63 | 12.40 | 1.06 | 27.90 | 73.37 |
| MedMamba | 50.83 | 6.91 | 5.89 | 8.94 | 92.73 |
| ConvKAN | 49.26 | 3.49 | 0.44 | 3.82 | 94.70 |
Key Findings¶
- Segmentation task: Leading models achieve strong DSC performance (up to 97.75%), yet NSD scores remain relatively low (up to 90.71%), indicating that precise boundary delineation remains the primary bottleneck. Significant DSC differences between BE and non-BE groups (p < 0.05–0.001) confirm that bone erosion negatively affects segmentation performance.
- Classification task: All models yield low BACC and F1 scores (maximum 52.64% and 12.40%, respectively), reflecting the extreme difficulty of the task. Models exhibit a pronounced bias toward the non-BE class (high specificity / low sensitivity), with severe class imbalance being the core challenge.
- Foundation model gap: SAM and MedSAM substantially underperform supervised models on carpal bone segmentation, demonstrating that general-purpose foundation models still have considerable room for improvement in fine-grained medical image segmentation.
Highlights & Insights¶
- RAM-W600 is the first publicly available carpal bone instance segmentation dataset, covering both segmentation and BE scoring tasks, thereby filling a critical data gap in RA wrist CAD research.
- Multi-center data sourcing (6 institutions) enhances data diversity and the reliability of generalization evaluation.
- Benchmark results clearly reveal the shortcomings of current models with respect to bone overlap, blurred boundaries, and erosion-induced deformation, providing concrete directions for future research.
- Mamba-based architectures (SwinUMamba, UMambaEnc) demonstrate a favorable balance between performance and parameter efficiency.
Limitations & Future Work¶
- RA cases are predominantly sourced from a single geographic region in Japan, resulting in relatively high demographic homogeneity that may limit model generalizability across different ethnicities and regions.
- The SvdH BE score distribution is severely imbalanced, with high-score cases being extremely scarce (scores of 3 and 5 are nearly absent), which impairs the training and evaluation of fine-grained scoring models.
- Only binary classification (presence vs. absence of BE) is evaluated; multi-grade regression scoring is not explored.
- Downstream tasks aligned with clinical workflows (e.g., JSN progression quantification, longitudinal monitoring) are not investigated.
Related Work & Insights¶
- This dataset can inspire multi-task learning frameworks that jointly address carpal bone segmentation and BE detection.
- The dataset construction methodology is extensible to RA assessment of other joints (fingers, feet).
- The extreme class imbalance in the BE classification task provides a valuable testbed for few-shot and imbalanced learning research in medical imaging.
Rating¶
- Novelty: ⭐⭐⭐⭐ First publicly available dataset of its kind, filling an important gap.
- Experimental Thoroughness: ⭐⭐⭐⭐ Broad benchmark coverage, though multi-grade scoring and downstream task validation are absent.
- Writing Quality: ⭐⭐⭐⭐ Clear structure with thorough statistical analysis.
- Value: ⭐⭐⭐⭐⭐ High practical utility for CAD research in the RA domain.