Harrison.Rad 1.5 Technical Report: A radiology foundation model that can draft reports from images, priors and clinical context

2026-07-08 · Source: cs.CV updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Health & Medical Research, Data Science & Analytics · Depth: Expert, extended

Summary

Harrison.Rad 1.5 (HR1.5) is a radiology-specific multimodal large language model designed to draft reports from images, prior studies, and clinical context, addressing growing imaging demand and radiologist shortages. Trained via a three-stage pipeline on approximately 6 million image-report instances, it covers plain-film radiology including chest, musculoskeletal, abdominal, spine, pelvic, and mammography. HR1.5 was evaluated using a Findings-Diagnosis scoring framework and benchmarked on RadBench and a simulated FRCR 2B Short Case examination. HR1.5+ achieved a 62% pass rate on the FRCR 2B Short Case, outperforming all other evaluated models. It also leads in closed-format clinical questions and mammography, with HR1.5+ scoring 77.8% on CBIS-DDSM. The report also details explainability features like Grad-CAM heatmaps and confidence estimation.

Key takeaway

For radiology department heads evaluating AI solutions to mitigate workforce shortages, Harrison.Rad 1.5 (HR1.5) offers a robust, specialized foundation model capable of drafting high-quality reports across diverse plain-film studies. You should consider HR1.5 for its demonstrated ability to meet FRCR examination standards and its advanced explainability features, which can streamline radiologist workflows and enhance diagnostic confidence. Its performance on complex cases and multi-finding reasoning suggests a significant reduction in reporting effort.

Key insights

Harrison.Rad 1.5 excels in radiology report generation by integrating multimodal data through a specialized multi-stage training pipeline.

Principles

• Domain adaptation improves model capacity.
• Hard negatives sharpen vision encoder representations.
• Clinically relevant data engineering is crucial.

Method

HR1.5 uses a three-stage pipeline: domain adaptation of a base LLM on radiology reports, contrastive vision-encoder training with curriculum-based hard negatives on ~6M image-report instances, and VQA fine-tuning on multi-turn clinical conversations.

In practice

• Use Findings-Diagnosis for clinical accuracy.
• Implement question-sensitive Grad-CAM for interpretability.
• Correct confidence scores for attention sinks.

Topics

• Radiology Foundation Models
• Multimodal AI
• Medical Report Generation
• Clinical Decision Support
• Explainable AI
• Diagnostic Imaging

Code references

• vllm-project/vllm
• harrison-ai/radbench

Best for: AI Scientist, Machine Learning Engineer, Research Scientist

Related on AIssential

• 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 multimodal LLM that drafts radiology reports, meeting FRCR standards through a specialized three-stage training.
• Vision Foundation Models in Radiology: A Scoping Review of Data, Methodology, Evaluation and Clinical Translation — Radiology-specific Vision Foundation Models show promising transferability but face significant hurdles for clinical translation.
• MARCH: Multi-Agent Radiology Clinical Hierarchy for CT Report Generation — A multi-agent framework mimicking radiology hierarchies improves CT report generation fidelity and accuracy.
• Scalable Training of Spatially Grounded 2D Vision-Language Models for Radiology — Spatially grounded radiology VLMs can be trained at scale without manual annotations using LLM-based curation.
• Beyond Scalar Scores: Exploring LLM-based Metrics for Clinical Significance Evaluation in Radiology Reports — LLM evaluators for radiology reports suffer "discrimination bias," but targeted data synthesis and fine-tuning can sharpen clinical significance boundaries.
• Spatio-Temporal and Clinical Conditioning for Fine-Grained Radiology Report Retrieval — Anatomical, temporal, and clinical conditioning significantly improves automated radiology report retrieval and generation, addressing current method limitations.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CV updates on arXiv.org.