Adjunctive artificial intelligence (AI) may facilitate reduced reading time and improved detection of chest X-ray abnormalities, ranging from pleural effusion and consolidation to lung nodules and pneumothorax, across radiologists of varying experience levels.
For the retrospective study, recently published in Radiology, researchers reviewed data from 500 patients (mean age of 54) who had chest X-rays and thoracic computed tomography (CT) within a 72-hour period.1 Twelve radiologists assessed half of the chest X-rays with the adjunctive AI software ChestView version 1.2.0 (Gleamer) and half of the X-rays without AI. The reviewing radiologists included four thoracic radiologists, four general radiologists and four radiology residents, according to the study.
The researchers found that adjunctive AI led to an average 26.2 percent increase in the sensitivity rate for detecting pneumothorax. Specifically, general radiologists had a 38.6 percent increase, radiology residents had a 34.3 percent increase and thoracic radiologists had a 5.7 percent increase.1
For pleural effusion, the use of adjunctive AI facilitated an average sensitivity increase of 8.5 percent. General radiologists had a 12.6 percent increase in sensitivity with AI, followed by an 8.1 percent increase for thoracic radiologists and a 5 percent increase for radiology residents.1
The study authors noted that the adjunctive AI software netted an average 14.1 percent increase in sensitivity for consolidation with a 21.1 percent increase for radiology residents, a 10.9 percent increase for general radiologists and a 10.3 percent increase for thoracic radiologists.1
“For critical conditions like pneumothorax, pleural effusion and consolidation, AI assistance led to notable sensitivity improvements, suggesting its potential to hasten detection and intervention,” wrote study co-author Guillaume Chassagnon, M.D., Ph.D, an associate professor in the Department of Thoracic Imaging at the Cochin Hospital in Paris, France, and colleagues.
Three Key Takeaways
- Improved sensitivity for abnormalities. The study found that the use of adjunctive AI, specifically ChestView version 1.2.0, led to a notable increase in sensitivity for detecting various chest X-ray abnormalities. There was an average increase of 26.2 percent in sensitivity for pneumothorax, 8.5 percent for pleural effusion, and 14.1 percent for consolidation. These improvements in sensitivity were observed across radiologists with varying levels of experience, indicating the potential of AI to enhance detection of critical conditions.
- Reduced reading time. The study demonstrated that the use of adjunctive AI software resulted in a significant reduction in reading time for chest X-rays. On average, there was a 25-second reduction, representing a 31 percent decrease in reading time. This reduction was more pronounced for radiology residents (30 percent), general radiologists (34 percent), and thoracic radiologists (27 percent). The time saved was particularly notable for radiographs without abnormalities, aligning with the majority of chest radiographs in clinical practice.
- Specificity improvements and considerations. The researchers reported average specificity increases with the use of adjunctive AI, including 3.9 percent for pleural effusion, 3.7 percent for mediastinal and hilar masses, and 2.9 percent for consolidation. However, there was no increase in specificity for pneumothorax.
The use of the adjunctive AI software also led to a mean 25 second reduction (31 percent) in reading time, according to the study authors. They noted a 27 percent reduction in reading time for thoracic radiologists, a 30 percent reduction for radiology residents and a 34 percent reduction for general radiologists.1
“As in the study by Shin et al, we observed that the time saved in reading is greater for radiographs without abnormalities, which represent the majority of chest radiographs in clinical practice,” added Chassagnon and colleagues.1,2
The researchers also noted average specificity increases of 3.9 percent for pleural effusion, 3.7 percent for mediastinal and hilar masses, and 2.9 percent for consolidation. There was no increase in specificity for pneumothorax, according to the study authors.1
(Editor’s note: For related content, see “FDA-Cleared AI Triage Software for Chest X-Rays Offers Enhanced Detection of Pleural Effusion and Pneumothorax,” “Study Shows Benefits of AI in Detecting Lung Cancer Risk in Non-Smokers” and “Can AI Improve Triage Efficiency in Radiology Workflows for Follow-Up X-Rays?”)
In regard to study limitations, the authors acknowledged that the reviewing radiologists were able to assess the X-rays without time constraints and were not aware of the clinical indication nor the patient’s medical history, conditions that may not reflect clinical practice. The researchers noted an average of 2.2 abnormalities per patient and that half of the patients had at least one visible abnormality on chest X-ray. They conceded that these findings don’t reflect clinical practice.
References
1. Bennani S, Regnard NE, Ventre J, et al. Using AI to improve radiologist performance in detection of abnormalities on chest radiographs. Radiology. 2023;309(3):e230860. doi: 10:1148/radiol.230860.
2. Shin HJ, Han K, Ryu L, Kim EK. The impact of artificial intelligence on the reading times of radiologists for chest radiographs. NPJ Digit Med. 2023;6(1):82.
