UB researchers developing tool to spot AI‑generated radiology reports

BUFFALO, N.Y. – Impersonating a doctor to create a fake radiology report. Inserting nonexistent fractures into real X-ray images. Both are ways in which artificial intelligence can be misused for insurance fraud, falsifying disability or malpractice claims, and committing other cybercrimes.

While not common, AI-generated medical reports have the potential to cause serious problems in the medical and insurance industries.

To combat this growing threat, a team of UB researchers has developed what they believe is the first AI system designed to distinguish between radiology reports written by humans and those generated by AI.

“With generative AI becoming more capable of producing remarkably convincing radiology reports, there’s a greater risk of fabricated reports being used to falsify medical histories and support fraudulent claims,” said lead research investigator Nalini Ratha, PhD, SUNY Empire Innovation Professor in the Department of Computer Science and Engineering at UB. “Radiology reports have highly specialized structure, vocabulary and stylistic norms, making general-purpose detectors unreliable. Therefore, our goal was to build a detection framework designed specifically for radiology that can distinguish clinician-written medical documentation from synthetic text before it reaches clinical or insurance workflows.”

The UB team – consisting of Ratha and PhD students Arjun Ramesh Kaushik and Tanvi Ranga – presented their study, “Detecting Synthetic Radiology Reports Using Style Disentanglement,” at the 2025 GenAI4Health workshop held during the Conference on Neural Information Processing Systems in San Diego in December.

A first-of-its-kind dataset

As part of their study, Ratha, Ranga and Kaushik built a dataset of 14,000 pairs of radiologist-authored and AI-generated chest X-ray reports. They used two approaches to generate the synthetic reports:

• Text-to-text: Paraphrasing real radiologist reports using advanced LLMs.
• Image-to-text: Generating full reports directly from chest radiographs using medical vision-language models (VLMs).

The dataset is the first to combine both text‑based and image‑based synthetic radiology reports, researchers say, marking a major step forward for trustworthy AI research in health care. All samples focused solely on the findings section, which is the portion of a report that captures the radiologist’s detailed analysis and includes extensive domain-specific terminology and descriptive language.

“The findings section is both central to authorship attribution and the one most susceptible to exploitation,” Ratha said.

Detection method shows high accuracy

The next phase of the study involved developing the authorship‑detection framework built to operate on this dataset. Although LLMs can replicate clinical terminology, they struggle to mirror the stylistic characteristics of human‑authored radiology reports. Recognizing this gap, the UB researchers created a BERT–Mamba–based detection model designed to separate each report’s stylistic features from its underlying clinical content.

“AI systems leave subtle stylistic fingerprints such as patterns in phrasing, punctuation and word choice that differ from how radiologists naturally write. By disentangling style from content and treating it as its own measurable feature, our model was able to detect those patterns with exceptional precision,” Kaushik said.

The UB model distinguished human‑written reports from synthetic ones with high accuracy and consistency, achieving Matthews correlation coefficient (MCC) scores between 92% to 100% in both text-to-text and image-to-text categories. MCC is a rigorous performance metric that evaluates all types of prediction outcomes and is especially valuable for testing models on imbalanced or complex datasets. Even when AI outputs closely resembled the original reports, text‑to‑text detection accuracy still exceeded 99%. The framework also held up in cross‑LLM tests, correctly spotting AI‑generated reports from models it had never seen before.

“What we found is LLMs tend to write in polished, expansive language, while clinicians write in concise, direct terms. Radiologists use straightforward terms like ‘heart’ or ‘lung.’ LLMs often replace them with more elaborate phrases like ‘pulmonary vasculature,’ which became a clear stylistic signal that our model learned to separate,” Ranga said.

Despite the impressive results, the team is continuing to refine the dataset and benchmark detection model as they prepare for the model for public release. Next steps include fine‑tuning VLMs to improve clinical alignment, and expanding the dataset to additional radiology categories while incorporating a wider range of AI models. Eventually, they envision the dataset serving as a controlled testbed for future studies aimed at rigorously evaluating the reliability, trustworthiness and safe deployment of AI‑generated clinical narratives.

The researchers also see AI systems, as they become more sophisticated and tailored to fields like radiology, as a tool that can save radiologists significant time and help them manage increasing workloads.

Applications beyond medicine

While his team’s research focused on radiology, Ratha says the implications of their study extend beyond health care. The same style‑based detection approach could also help safeguard industries that are increasingly vulnerable to AI‑generated forgeries, fabricated records and synthetic narratives, including insurance, finance, journalism, education and the legal profession.

“People are using AI to generate quick answers, whether it be a medical report, a student essay or a legal brief. The consequences are minor in some settings, but in mission-critical applications such as health care or law, the stakes are far higher,” Ratha said. “Any field that relies on written records could benefit from a detection system that identifies whether a document’s style aligns with human authorship, even when the content appears legitimate.”