← News

Models & toolsSources: arXiv · 20 Nov 2026

Rad-Linter pairs chest X-ray images with reports to flag possible errors

Researchers describe a small fine-tuned language model that checks radiology reports against image-derived pathology scores and outperformed a rule-based baseline in controlled tests.

In plain words

Rad-Linter is a tool that reads a chest X-ray report together with scores a separate image classifier produced for nine conditions, and flags cases where the words and the image evidence do not match. The language model was fine-tuned, meaning adjusted on a small set of examples, to judge this consistency. In the reported tests it made fewer mistakes than a simpler keyword-based checker.

What was announced

The system, described in an arXiv paper, uses a pretrained image classifier that assigns scores for nine pathologies, while a 1.5B-parameter language model fine-tuned with LoRA on 1,816 examples judges whether those scores agree with the report text. The authors compared it with a keyword-and-negation rule baseline that received the same reports and scores.

Evaluation relied on controlled pairings from the OpenI dataset, where each test item combines two real reports with two real score vectors from other patients, so neither input alone can exceed 50% accuracy. On 131 such quads, the paper reports 98.5% pair accuracy for Rad-Linter versus 80.5% for the rule baseline, with all four pairs correct in 94.4% versus 56.5% of quads. On a second set of 181 quads including reports where automatic labelers disagree, pair accuracy was 93.4% versus 75.3%. The advantage was concentrated on items requiring interpretation of how findings are worded; the two checkers performed similarly on omission items, and the same model without fine-tuning stayed near 50% accuracy.

What it does not show

  • The evidence is the paper's abstract only; full methods, datasets and peer review status are not visible, and the results come from controlled counterfactual pairings rather than described deployment.
  • Consistency labels were derived from automatic report labels and score thresholds, so label quality itself may affect the reported accuracy figures.

For security and IT teams

Scope
The reported evaluation covers chest X-ray reports checked against nine pathology scores from one classifier, not general radiology reporting.

Sources

Source-support review completed 21 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 21 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.