Harrison.Rad 1.5 Technical Report: A radiology foundation model that can draft reports from images, priors and clinical context
Harrison.Rad 1.5 is a radiology-specific multimodal large language model trained through a three-stage pipeline to generate structured and unstructured reports from diverse x-ray and mammography images, clinical context, and prior studies, achieving state-of-the-art performance on clinical benchmarks including the simulated FRCR examination.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
The Big Problem: Too Many Scans, Not Enough Eyes
Imagine a hospital where the number of X-rays and scans is growing like a rapidly expanding balloon, but the number of radiologists (the doctors who read them) is only growing like a slow, steady snail. This creates a massive traffic jam of unreported scans.
The paper argues that simply hiring more doctors or training them faster won't fix this. The real solution is to give the existing doctors a "super-assistant" that does the heavy lifting of writing the report, so the doctor just has to check it.
The Solution: Harrison.Rad 1.5 (HR1.5)
Think of HR1.5 not as a simple calculator that counts bones, but as a highly specialized medical intern who has read millions of X-rays and written millions of reports.
- What it does: It looks at an X-ray, reads the patient's history (like "patient has a cough"), checks their old X-rays to see what's changed, and then drafts a full medical report.
- What it covers: It's not just for chest X-rays. It handles X-rays of the spine, hips, knees, abdomen, and even mammograms. It's a "generalist" for flat X-rays, not just a specialist for one body part.
How It Was Trained: The Three-Step Boot Camp
The paper describes a three-stage training process, which is like how you might train a new employee:
- Reading the Manuals (Domain Adaptation): First, the AI was fed a massive library of real radiology reports so it learned the specific language doctors use. It learned that "opacity" isn't just a blurry spot; it's a specific medical term.
- The "Spot the Difference" Game (Contrastive Learning): The AI was shown millions of pairs of images and reports. Crucially, it was shown "hard negatives"—images that look almost identical but have a tiny, critical difference (like a tiny fracture vs. no fracture). This taught the AI to pay attention to the clinically important details, not just the general shape of the image.
- Role-Playing with Doctors (Instruction Tuning): Finally, the AI practiced having conversations with doctors. It learned to answer specific questions ("Is there a fracture?") and draft reports in the exact style hospitals expect.
The Hardware Upgrade: The team upgraded the "brain" of the AI to run on the newest, most powerful computer chips (NVIDIA B200 class), allowing it to process data much faster and more efficiently than previous versions.
The Big Test: The "Medical Board Exam"
To prove this AI is actually good, the researchers didn't just ask it to guess; they gave it a mock version of the Royal College of Radiologists (FRCR) exam. This is the real, difficult test human doctors must pass to become specialists.
- The Result: HR1.5 was the only system (including other famous AI models and general chatbots) to pass the exam.
- The Comparison: While general AI models (like the ones you might chat with online) scored below 50% (failing), HR1.5 scored high enough to pass. It even outperformed the previous version of itself (HR1) and other medical-specific AIs.
Why Standard Tests Fail (The "Word Match" Trap)
The paper points out a funny flaw in how we usually test AI. Most tests check if the AI's answer uses the same words as the correct answer (like a teacher grading a spelling test).
- The Problem: An AI could write a perfect medical report but use slightly different words, and a standard test would mark it wrong. Or, it could write a nonsense report that happens to use the right keywords, and the test would mark it right.
- The Fix: The researchers created a new scoring system called "Findings-Diagnosis." Instead of checking for word matches, it checks for medical truth. Did the AI spot the pneumonia? Did it correctly say the heart size is normal? Did it avoid making up diseases that aren't there? This system is much stricter and more honest about clinical accuracy.
How It Thinks: "Explainability"
One of the biggest worries with AI is that it's a "black box"—it gives an answer, but you don't know why. The paper shows how HR1.5 tries to be transparent:
- Heatmaps: When the AI says, "There is a fracture here," it can highlight the exact spot on the X-ray that made it say that. It's like the AI pointing its finger at the evidence.
- Confidence Meter: The AI can tell you how sure it is. If it's looking at a weird, old, or unclear image, it will say, "I'm not 100% sure about this."
- Checking for Hallucinations: The system has a built-in "lie detector." If the AI is just guessing because it's confused, the system can detect that its internal signals are shaky and lower its confidence score.
The "Apple in a Chest" Test (Out-of-Distribution)
The researchers tested what happens if you show the AI something it has never seen before. They showed it a fake X-ray of an apple inside a human chest.
- The Result: The AI didn't say, "I don't know what that is." Instead, it confidently said, "That looks like a breast implant."
- The Lesson: This shows a limitation. If the AI sees something totally new, it tries to fit it into a box it already knows. It's not perfect at recognizing "unknown unknowns," which is why human doctors are still needed to double-check.
The Bottom Line
Harrison.Rad 1.5 is a specialized AI tool designed to help radiologists write reports faster and more accurately. It has passed a simulated specialist exam that other AIs failed. It doesn't replace the doctor; it acts as a draft-writer that the doctor reviews.
Important Note: The paper explicitly states that this tool is not yet approved for clinical use. It is currently a research prototype. It is not a medical device, and you cannot use it to diagnose patients yet. It is being released so scientists can study it, improve it, and figure out how to make it safe for the real world.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.