aiqbench analysis
This research introduces QureRadEmbed, a specialized 4B parameter encoder that significantly improves the retrieval and comparison of radiology reports by explicitly modeling fine-grained attributes like laterality and severity. By combining deterministic ranking targets with reasoning-aligned supervision, the model achieves near-perfect ordering accuracy on synthetic attributes and offers a substantial speed advantage over existing generative evaluators for whole-report comparisons. The work demonstrates that structured embeddings can effectively capture complex disease relationships and subtle discrepancies in medical imaging contexts without relying solely on broad biomedical similarity metrics.
QureRadEmbed: Structuring Radiological Similarity through Attribute and Reasoning Supervision
arXiv:2609.33075v1 Announce Type: new Abstract: Radiological similarity depends on disease relationships and on fine details such as laterality, lobe, severity, size, and certainty. Broad biomedical similarity can overlook these qualifiers, particularly when several attributes vary together. We introduce QureRadEmbed, a 4B radiology-aware encoder trained with two complementary signals: RadSim supplies deterministic, attribute-decomposed ranking targets, while RadThought aligns reports with hierarchical evidence and reasoning descriptions. A three-stage curriculum combines these signals with report triplets, finding perturbations, and single- and cross-attribute contrasts. The final model achieves 0.996 mean ordering accuracy across ten controlled synthetic attributes and raises Spearman correlation with the designed joint-attribute targets from 0.501 to 0.976. On external findings-to-impression retrieval, Recall@1 reaches 10.5% on Open-I, 12.4% on testing XR, and 42.9% on testing CT, compared with 6.6%, 5.7%, and 31.4% for its backbone. Frozen embeddings support finding extraction with only 100 labeled testing-XR reports (macro-F1 0.481 versus 0.412 for the backbone). Whole-report comparison costs 8.8 seconds per 1,000 pairs in our benchmark, versus 2,755.1 seconds for the generative evaluator GREEN. Sentence-level comparison improves sensitivity to local discrepancies, although generative evaluation remains stronger on several expert-rated and subtle-error tasks. The results support reusable radiology-aware representations for search, structured report indexing, and efficient report comparison.
Abstract:Radiological similarity depends on disease relationships and on fine details such as laterality, lobe, severity, size, and certainty. Broad biomedical similarity can overlook these qualifiers, particularly when several attributes vary together. We introduce QureRadEmbed, a 4B radiology-aware encoder trained with two complementary signals: RadSim supplies deterministic, attribute-decomposed ranking targets, while RadThought aligns reports with hierarchical evidence and reasoning descriptions. A three-stage curriculum combines these signals with report triplets, finding perturbations, and single- and cross-attribute contrasts. The final model achieves 0.996 mean ordering accuracy across ten controlled synthetic attributes and raises Spearman correlation with the designed joint-attribute targets from 0.501 to 0.976. On external findings-to-impression retrieval, Recall@1 reaches 10.5% on Open-I, 12.4% on testing XR, and 42.9% on testing CT, compared with 6.6%, 5.7%, and 31.4% for its backbone. Frozen embeddings support finding extraction with only 100 labeled testing-XR reports (macro-F1 0.481 versus 0.412 for the backbone). Whole-report comparison costs 8.8 seconds per 1,000 pairs in our benchmark, versus 2,755.1 seconds for the generative evaluator GREEN. Sentence-level comparison improves sensitivity to local discrepancies, although generative evaluation remains stronger on several expert-rated and subtle-error tasks. The results support reusable radiology-aware representations for search, structured report indexing, and efficient report comparison.
39 pages, 10 figures, including appendices with per-tag labeling results. Janhavi Prabhu and Sahil are co-first authors. Manoj Tadepalli is the corresponding author
Artificial Intelligence (cs.AI)
arXiv:2609.33075 [cs.AI]
(or arXiv:2609.33075v1 [cs.AI] for this version)
https://doi.org/10.48550/arXiv.2609.33075 arXiv-issued DOI via DataCite (pending registration)
Full text under the publisher's open licence · 239 words
Related coverage
Biomedical Imaging's Real Bottleneck Is the Data, Not the Model
Databricks · Oct 8, 2026
[AINews] not much happened today
Latent Space · Sep 10, 2026
J++ Lens: Jacobian Filtering Enables More Faithful Workspace Lenses
LessWrong · Oct 8, 2026
Introducing CARE-X: Towards Clinically Useful Radiology VLMs with Auxiliary Supervision, Reward-Aligned Learning, and Tool-Augmented Measurement
Microsoft Research · Aug 11, 2026