Opening — Why this matters now
Radiology AI is not short on models. It is short on structure.
Over the past decade, thousands of deep learning systems for lesion detection, segmentation, report drafting and generative enhancement have appeared across journals, conferences and preprints. The problem is no longer innovation velocity — it is navigability. Models are scattered across supplementary PDFs, personal GitHub accounts, institutional pages and occasionally, abandoned repositories.
For a clinician or AI team trying to implement a real system, this fragmentation translates into friction: missing weights, broken links, undocumented validation strategies, ambiguous intended use.
OpenRad attempts to solve this structural bottleneck. And in doing so, it quietly surfaces a deeper question: what does it actually take to move radiology AI from publication to practice?
Background — The Reporting Gap No One Owns
Radiology has not ignored governance. The RSNA AI Roadmap and RadLex ontology introduced structured reporting standards. CLAIM guidelines improved methodological transparency. Yet adherence remains inconsistent.
The result is a paradox:
| Dimension | Research Output | Practical Usability |
|---|---|---|
| Publication volume | Extremely high | High noise |
| Metadata structure | Inconsistent | Hard to compare |
| Code availability | Variable | Often unusable |
| Pretrained weights | Rare | Limits adoption |
Existing repositories either curate a limited subset of model cards or include entries without executable code. Many catalog papers; few verify functionality.
OpenRad addresses this gap not by publishing new models — but by systematizing the ecosystem.
That distinction matters.
Analysis — What OpenRad Actually Does
OpenRad aggregates 1,694 open-access radiology AI models extracted from 5,239 screened records across PubMed, arXiv and Scopus.
The pipeline includes:
-
Automated metadata extraction using a locally hosted open-source LLM (gpt-oss:120b).
-
Structured population of RSNA AI Roadmap JSON schema fields.
-
Automated interrogation of GitHub repositories to verify:
- Presence of pretrained weights
- Availability of interactive demos
-
Manual review by ten AI researchers.
The system evaluates extraction stability across 225 duplicate generations using lexical similarity metrics.
Stability Snapshot
| Field | Levenshtein Ratio (Mean) | Interpretation |
|---|---|---|
| Title | 100% | Deterministic |
| Authors | >90% | High consistency |
| Repository Link | ~90% | Stable |
| Architecture | ~67% | Moderate variance |
| Limitations | ~50% | High rephrasing |
Key insight: structured fields remain stable; narrative sections vary lexically but retain semantic meaning.
Manual review classified 78.5% of corrections as minor — indicating that automated large-scale extraction is viable when paired with expert oversight.
In other words: LLMs can scale metadata curation, but humans still anchor fidelity.
Findings — What the Data Reveals About Radiology AI
Beyond aggregation, OpenRad functions as a live analytics layer over the field.
1. Architecture Distribution
| Architecture Type | Share |
|---|---|
| Transformers | 33.4% |
| CNN Variants | 33.3% |
| Diffusion Models | 12.1% |
| U-Net Variants | 10.1% |
| GANs | 8.1% |
| YOLO / LSTM | <2% |
Transformers have matched CNNs in dominance — a structural shift from earlier imaging paradigms.
2. Geographic Concentration
| Country | Approx. Publications |
|---|---|
| China | ~400 |
| United States | ~300 |
| Secondary Tier (EU, India, East Asia) | 50–150 each |
Radiology AI research is geographically concentrated, mirroring broader AI power dynamics.
3. Modality-Specialty Focus
MRI dominates neuroradiology (621 studies). Chest imaging (CT and X-ray) follows. Nuclear medicine and PET remain underrepresented.
4. Evaluation Metrics by Task
- Classification: Accuracy (229), AUC (166), F1-score (112)
- Segmentation: Dice Similarity Coefficient (308 dominant)
- Generative: SSIM (88), PSNR (79)
The evaluation culture remains task-specific but heavily metric-driven.
5. Recurring Limitation Themes
Most common limitation keywords:
- “external validation”
- “single center”
- “small dataset”
- “generalization”
Translation gap identified: high internal performance, weak multi-site validation.
Implications — Why This Is Bigger Than a Repository
OpenRad is not merely a directory.
It represents a governance layer for open radiology AI.
Three strategic implications emerge:
1. Discoverability as Infrastructure
AI adoption depends less on model novelty and more on structured accessibility. Repositories that verify weights and demos reduce integration friction dramatically.
2. LLMs as Meta-Research Tools
The use of an open-source LLM for schema-constrained extraction signals a scalable method for maintaining technical registries.
This approach could extend beyond radiology — into drug discovery models, surgical robotics AI, pathology systems.
3. Transparency Pressure on the Field
By verifying actual repository health (weights, links, demos), OpenRad implicitly pressures authors toward reproducibility.
Public accountability changes behavior.
Limitations — Where Structure Still Ends
- Extraction depends on access to full texts; paywalled papers reduce fidelity.
- Initial implementation does not populate every RSNA field.
- Sustainability depends on community participation.
In other words: structure exists, but maintenance determines longevity.
Conclusion — From Model Proliferation to Model Governance
Radiology AI has entered its second phase.
Phase one was invention.
Phase two is organization.
OpenRad demonstrates that scalable curation, LLM-assisted metadata extraction and expert validation can transform scattered innovation into a coherent, searchable system.
In a field where clinical translation depends on traceability, reproducibility and clarity of intent, that shift may matter more than the next marginal performance gain.
Structure is not glamorous.
But structure is what scales.
Cognaptus: Automate the Present, Incubate the Future.