A RECENT study has highlighted the potential of AI to improve radiology workflows by automating the identification of unremarkable chest radiographs. As radiology departments face increasing workloads, AI could help reduce case volumes by automatically excluding normal findings, allowing radiologists to focus on more complex cases without compromising diagnostic accuracy.
The retrospective study analysed 1,961 chest X-rays from adult patients across four Danish hospitals between 1–12 January 2020. The AI tool was tested at various sensitivity levels to assess its ability to distinguish unremarkable chest radiographs from those with potential pathologies following which, two thoracic radiologists, blinded to the AI results, classified the radiographs as either “remarkable” or “unremarkable.” The AI then assigned a probability score for the remarkableness of each X-ray, and its specificity was measured at sensitivity thresholds of 99.9%, 99.0%, and 98.0%.
At these sensitivity levels, the AI correctly excluded pathology in 24.5%, 47.1%, and 52.7% of unremarkable chest X-rays, respectively. These results translate into potential reductions in case volumes of 9.1%, 17.5%, and 19.6%, offering a significant opportunity to streamline radiology workflows. When compared to the radiologists’ sensitivity of 87.2%, the AI demonstrated a critical miss rate of 2.2%, slightly higher than the radiologists’ 1.1%, however, the AI’s clinically significant misses were similar to the radiologists’ (4.1% vs. 3.6%). Importantly, when the AI’s sensitivity was increased to 95.4%, its performance in detecting critical misses matched or exceeded that of radiologists.
The study concluded that this AI tool has the potential to exclude pathology in up to 52.7% of unremarkable chest radiographs while maintaining low rates of diagnostic errors. Although the results are promising, the researchers recommended further prospective studies to validate the AI’s impact in real-world clinical settings. If proven effective, this technology could revolutionise radiology by reducing workload without sacrificing diagnostic safety.
Katie Wright, EMJ
Reference
Plesner LL et al. Using AI to identify unremarkable chest radiographs for automatic reporting. Radiology. 2024;312(2):e240272.
