Traditional keyword searches in radiology information systems may soon be a thing of the past. Researchers have introduced RadSearch, a domain-specific semantic search model designed to retrieve radiology reports with unprecedented precision, while also enhancing large language model diagnostic performance.

RadSearch was trained using a scalable method that bypasses the need for extensive manual labeling by researchers at the University of Alabama at Birmingham and the University of California, San Francisco. By leveraging contrastive learning on structured radiology reports, RadSearch captures the semantic relationships between imaging findings and impressions more accurately than existing models.

In tests using over 13,000 reports from multiple institutions, RadSearch retrieved relevant findings in 83% for free-text clinical queries, outperforming the current state-of-the-art model GTE-large, which achieved 65.7%. It also matched the correct location for findings in 89.8% cases compared to GTE-large's 58.8%

Boosting AI Diagnosis

RadSearch also demonstrated its clinical value by improving the diagnostic performance of a leading large language mode (LLM), Llama 3.1 8B Instruct. Without search assistance, the LLM correctly diagnosed only 30% of cases based on report findings alone. With RadSearch integration, diagnostic accuracy jumped to 61%, significantly outperforming the LLM's performance when assisted by GTE-large (47%).

The research shows semantic search models like RadSearch can streamline radiology workflows and improve the accuracy of AI-driven decision support, according to the study authors.

How RadSearch Works

RadSearch uses a Siamese neural network architecture initialized with weights from RadBERT-RoBERTa-4m, optimized for semantic similarity between report sections. The model was trained to distinguish between matching findings-impression pairs and mismatches without requiring tedious manual annotations.

The tool offers two main functionalities for radiologists:

• Finding similar cases: By inputting a description of an unfamiliar finding, radiologists can retrieve reports with similar findings even when exact keywords do not match.

• Diagnosis suggestion: Integrated with an LLM, RadSearch can provide differential diagnoses based on retrieved cases, accelerating clinical decision-making.

Clinical and Research Impact

RadSearch could dramatically enhance radiologists' ability to find precedent cases, improve educational tools for residents, and reduce the manual effort required to curate research datasets, noted researchers. Its superior performance highlights the importance of domain-specific semantic models over general AI search solutions.

The researchers envision future integration of RadSearch into clinical PACS and RIS systems, with further testing in live clinical environments.

Disclosures can be found in the published study.

Source: Radiology