Inference-time Correction of Errors in AI-Generated Chest X-ray Radiology Reports

Raziuddin Mahmood, Tanveer F Syeda-Mahmood

Keywords: Fact-checking, regressor models, Chest X-ray, radiology reports

Abstract: Automated radiology report generators are being increasingly explored in clinical workflow pilots, particularly for chest X-ray imaging. However, their factual correctness with respect to the description of the findings has often been less than accurate, making their adoption slow and requiring detailed verification by clinical experts. In this paper, we propose an automatic report correction method that uses both image and textual information in automated radiology reports to spot identity and location errors in findings through fact-checking models. Prompts for a pre-trained large language model are then generated from the analysis of these errors to produce corrected sentences by selectively modifying target findings described in the automated report sentences. We show that this method of report correction, on the average, improves the report quality between 17-30% across various SOTA report generators over multi-institutional chest X-ray datasets.

Primary Area: other topics in machine learning (i.e., none of the above)

Submission Number: 23406