Implementation of Artificial Intelligence–Assisted Chest X-ray Interpretation: It Is About Time
S. Bodduluri,Sundaresh Ram
DOI: https://doi.org/10.1513/AnnalsATS.202303-195ED
2023-05-01
Abstract:Chest radiography (CXR) for the diagnosis of pulmonary complications offers portability and wide accessibility and is inexpensive. CXR is often the first imaging modality for the detection of lung disease and, thus, shapes disease management strategy (1). Despite the significance and obvious advantages, the inconsistent interpretation of CXR and associated reporting delays continue to be a significant burden in various healthcare settings (2–5). Alternatively, recent advances in convolutional neural networks have led to the birth of several stand-alone artificial intelligence (AI) tools to interpret CXR. Several studies have evaluated the impact of AI as an assistant to the workflow of radiologists in the detection and localization of key radiographic findings and in clinical decision making (6, 7). There is an augmented curiosity but a cautious adoption of AI-based diagnostics by radiologists (8). AI algorithms have been developed and deployed for specific CXR diagnostic tasks, such as detection of lung nodules (7), emphysema (9), pneumonia (10), and pneumothorax (11), which have exhibited radiologist-level performance and accuracy. Even though AI applications in CXR diagnosis have been shown to aid radiologists in detection of key radiographic findings, the benefit of AI-assisted CXR diagnosis and interpretation by nonradiologist physicians has not been well examined. Furthermore, it is unclear how deep learning and AI-based applications will integrate into a nonradiologist’s clinical workflow and what impact they will have on their efficiency. In this issue ofAnnalsATS, Lee and colleagues (pp. 660–667) designed a prospective, randomized, multicenter study to evaluate the impact of AI assistance on the detection and localization of lung lesions on CXR by nonradiologist physicians (12). The CXRs were randomized into an intervention group of 162 patients, in which CXR interpretation was aided by AI, and a control group of 161 patients, in which CXR interpretation was performed by nonradiologist physicians without the aid of AI. The accuracy of the AI algorithm was verified by three expert radiologists. In diagnosing the presence of lesions, the area under the curve (AUC) was 0.84 (95% confidence interval [CI], 0.77–0.90) for the intervention group, compared with 0.71 (95% CI, 0.64–0.79) for the control group (P value=0.017). The intervention was associated with significantly higher sensitivity and negative predictive value. In localizing lesions on CXR, the results followed a similar pattern, with an AUC of 0.80 (95% CI, 0.74–0.86) for physicians with AI assistance and 0.67 (95% CI, 0.60–0.75) in the control group (P=0.01). The study also reported between-group comparisons of the diagnostic accuracy when aided by chest computed tomography (CT) findings instead of AI. Interestingly, there was no significant difference between the intervention and the control groups in detecting lung lesions. At regional localization of lesions based on chest CT findings, the intervention group had a significantly higher area under the precisionrecall curve of 0.86, versus 0.69 in the control group (P=0.009). The study also included the effect of AI assistance on physicians’ clinical decisions as a secondary endpoint. There were no significant differences between the intervention group and the control group in the follow-up recommendations of clinical procedures, including chest CT, bronchoscopy, biopsy, and surgery. This is a carefully constructed and much-needed study measuring the impact of AI assistance on nonradiologist physicians from several outpatient clinics. The real clinical benefits of any AI-based diagnostic platform should include higher diagnostic accuracy with improved reporting efficiency, resulting in reduced false-positive and false-negative rates. AI-aided radiologic interpretation of lung lesions provided a significant advantage for physicians in the detection and localization of lung lesions and decreased the false referral rate. The control group without AI assistance had a higher frequency of followup chest CT scans than the intervention group. We should, however, note certain limitations. Although AI helped in imagebased detection and localization, it is critical to note that there was no significant difference between the two groups in terms of follow-up clinical decisions. This may depend on the population selected and frequency of clinical decision making using CXR. The authors missed an opportunity by not evaluating the time saved and improvements to clinical workflows in the AI-assisted physicians group. Although AI results were verified by three expert radiologists, there is less clarity on the
Medicine,Computer Science