RAM-H1200: A Unified Evaluation and Dataset on Hand Radiographs for Rheumatoid Arthritis

Technical Summary of RAM-H1200: A Unified Evaluation and Dataset on Hand Radiographs for Rheumatoid Arthritis

Problem Statement
Rheumatoid arthritis (RA) assessment from hand radiographs necessitates a multi-level analysis that integrates anatomical structure modeling, fine-grained local pathological changes (specifically bone erosion, BE, and joint space narrowing, JSN), and clinically standardized scoring (Sharp/van der Heijde, SvdH). Existing public resources fail to support this unified paradigm. Large-scale datasets like MURA or RSNA Bone Age provide only global labels or abnormality detection without fine-grained structural masks. RA-specific datasets often lack pixel-level annotations, are restricted to the wrist region (e.g., RAM-W600), or focus on trauma rather than RA. Furthermore, quantitative analysis of bone erosion remains scarce, with most resources relying on coarse categorical grading rather than explicit spatial supervision for lesion extent and morphology. This lack of comprehensive data hinders the development of models capable of jointly capturing anatomical context, localized pathology, and clinical severity.

Methodology and Dataset Construction
The authors introduce RAM-H1200, a multi-center evaluation and dataset comprising 1,200 high-resolution hand posteroanterior (PA) radiographs collected from six medical centers in Japan. The dataset includes 241 RA patients and 50 non-RA controls, covering diverse imaging conditions and patient demographics.

The dataset is distinguished by its multi-level annotation strategy, performed by a multidisciplinary team including radiologists and orthopedic doctors:

1. Hand Bone Structure Instance Segmentation: Precise contour delineation for 30 anatomically defined structures across the entire hand, including phalanges, metacarpals, carpal bones, radius, ulna, and sesamoid bones, along with surrounding soft tissue.
2. Pixel-Level Bone Erosion (BE) Masks: Annotations for BE regions categorized by diagnostic certainty: high-confidence SvdH BE, moderate-confidence SvdH BE, and non-SvdH BE. This enables quantitative analysis beyond coarse grading.
3. SvdH-Defined Joint Regions of Interest (ROIs): Anatomical locations for BE (16 joints) and JSN (15 joints) scoring, following the SvdH protocol.
4. Joint-Level SvdH Scoring: Ordinal severity scores for both BE and JSN assigned at the joint level.

The dataset is split at the patient level to prevent data leakage and is accompanied by standardized preprocessing (resampling to 0.175 mm/pixel, left-hand flipping).

Key Contributions
The paper claims three primary contributions:

• A Unified Multi-Task Benchmark: RAM-H1200 is the first large-scale dataset to systematically support three hierarchically organized aspects of RA analysis: (i) quantitative BE analysis via pixel-level modeling, (ii) whole-hand bone structure modeling via instance segmentation, and (iii) standardized SvdH-based scoring for both BE and JSN. This design connects lesion-level quantification, structure-aware modeling, and clinical assessment.
• Fine-Grained and Anatomically Consistent Annotations: The dataset provides high-quality pixel-level annotations for both anatomical structures and erosion regions, preserving anatomical consistency to support local lesion analysis and global structural modeling.
• Comprehensive Benchmarking Framework: The authors establish standardized benchmarks with unified evaluation metrics for each sub-task, enabling systematic evaluation and cross-model comparison of intermediate task models, from anatomical modeling to lesion quantification and clinical scoring.

Experimental Results
The authors evaluated various state-of-the-art models (including U-Net variants, Transformers, Mamba-based architectures, and foundation models like SAM) on the test set.

• Hand Bone Structure Segmentation: Supervised models achieved strong performance, with Dice Similarity Coefficients (DSC) exceeding 96% for most architectures. SwinUMamba achieved the best overall results. However, performance degraded significantly in anatomically overlapping regions (e.g., the wrist), where projection-induced ambiguity obscures boundaries. Foundation models (SAM, MedSAM) performed substantially worse than supervised models, particularly in overlap regions.
• Hand BE Segmentation: This task remained a major open challenge. All models achieved relatively low DSC scores (below 20%), reflecting the difficulty of detecting small, sparse, and low-contrast lesions. SwinUMamba performed best (DSC ~19.59%), but models struggled with the trade-off between sensitivity and false positives, often missing small erosions or producing fragmented masks.
• SvdH Scoring (BE and JSN):
  • BE Scoring: Models showed limited ordinal agreement (QWK ~0.45 for MedMamba) and balanced accuracy. Performance was heavily skewed toward low-grade predictions due to class imbalance. Confusion matrices indicated that higher-grade erosions were rarely predicted correctly.
  • JSN Scoring: Performance was relatively better than BE scoring (QWK ~0.60 for MobileViT) but still fell short of clinical applicability. Models captured coarse patterns but struggled with fine-grained severity classification, particularly for mild narrowing (grades 1–2).

Significance and Claims
The paper posits that RAM-H1200 provides a single benchmark for comprehensive RA analysis, bridging the gap between anatomical structure modeling and clinically grounded scoring. The authors highlight that while anatomical modeling (bone segmentation) is substantially more mature, quantitative BE analysis remains a significant bottleneck. The results demonstrate that current models can effectively model global hand skeletons but struggle with the fine-grained, ambiguous, and projection-dependent nature of erosive lesions and joint scoring.

The authors emphasize that the dataset enables the first explicit formulation and benchmarking of quantitative BE analysis at scale, moving beyond coarse severity grading. They conclude that reliable RA assessment requires future research to incorporate structural priors, multi-scale modeling, and uncertainty-aware learning to address the intrinsic difficulties of RA pathology and radiographic imaging. The dataset is intended to facilitate the development of computer-aided diagnosis systems as assistive tools, rather than standalone diagnostic replacements.