Radiology-specific foundation model

Original Article ↗ Hacker News Discussion ↗

A new radiology-focused foundation model from Harrison.ai claims near–expert-level performance on a key UK radiology exam and outperforms general multimodal models like GPT‑4o on X‑ray benchmarks, sparking interest in its potential to ease radiologist workloads and reduce burnout. Commenters probe how robust these results are outside controlled tests, note the lack of open access to the model and its proprietary training data, and question whether such tools should be made directly available to patients for self-diagnosis. Many see the biggest near-term impact not in replacing radiologists, but in automating reporting, integrating with hospital systems, and prioritizing scans that truly need human review.

Overall reception

  • Many commenters are impressed by radiology-specific performance, especially on formal exams and radiograph tasks.
  • Others reserve judgment until third‑party validation and real‑world deployment data are available.
  • Some note this is far from the first “AI for radiology” effort; the commercial path is seen as the real difficulty.

Benchmarks, exams, and claims

  • Reported results on the FRCR 2B Rapids mock exam (radiographs only) are seen as strong; some ask if the model was trained on exam questions.
  • A developer states the model was not trained on FRCR questions and clarifies it only took a mock rapid‑reporting component, not the full exam, and that this has since been clarified on the site.
  • Comparisons to general multimodal LLMs (GPT‑4o, Gemini, Claude, etc.) are questioned because most weren’t trained specifically on diagnostic imaging.

Access, openness, and datasets

  • Several people can’t find a public model or code; access appears gated via a waitlist and eventual commercialization.
  • Commenters wish for an open, “LLaMA‑style” radiology foundation model; current open efforts are mostly narrow (e.g., lung cancer) or small research models.
  • Suggestions for datasets include TCGA/NCIA, DeepLesion, MIMIC CXR, and commercial vendors.

Clinical context and workflow

  • Radiologists stress that diagnosis is not pure image classification; patient demographics, history, and symptoms matter.
  • The model’s use of both images and chart data is praised as more realistic.
  • Multiple comments complain bitterly about RIS/PACS/EMR fragmentation and messy metadata; integration and data quality are seen as a larger barrier than model accuracy.
  • There is strong interest in tools that:
    • Auto‑structure dictations and reports.
    • Explain specific image regions with literature references.
    • Triage studies and reduce radiologist burnout.

Ethics, public access, and economics

  • One camp argues that restricting public access is largely greed; another cites safety, misdiagnosis, and system strain from “confidently wrong” self‑diagnosers.
  • Debate over whether access limits are “infantilization” vs necessary stewardship, with examples from different countries’ pharma and diagnostic access.
  • Concerns that AI may become another billable line item while justifying staff cuts and shifting liability.
  • Radiology remains highly competitive as a specialty; several argue AI is more likely to augment overworked radiologists than replace them soon.