Application of artificial intelligence in chest radiograph interpretation: Is the time ripe yet?

Artificial intelligence (AI) can be defined as technology that mimics human cognitive processes, such as learning, reasoning and problem-solving. AI applications in radiology are driven by the idea that medical images are a set of data that can be computed by a machine to extract useful information.1 Machine learning (ML) is a branch of AI that applies concepts and tools to build algorithms aimed at the automated detection of meaningful patterns in data. While radiologists mainly evaluate qualitative features, comparing them to a subjective reference standard, ML features are low-level properties that can be computed or measured easily. There are handcrafted features (manually defined by data scientists) and/or automatically extracted features (usually through deep-learning algorithms). Hence, AI-powered assessment could make a significant contribution in this field, reducing the variability in image interpretation and improving diagnostic accuracy.2 Chest radiograph (chest X-ray CXR), is the most commonly used first-line investigative technique for disease evaluation, due to its widespread availability, low costs and the possibility to be performed at the patient’s bed. A correct and rapid CXR report is decisive in choosing the proper treatment and improving the patients’ outcomes. Although CXR reading is considered a basic radiological skill, it remains challenging and depends on the radiologist’s experience, workload and environment.1 Most clinicians in outpatient clinics or the emergency room (ER) frequently interpret CXRs. Human error, reader inexperience, fatigue and interruptions contribute to interpretation inaccuracy, and the availability of experienced thoracic radiologists is limited.3 Due to this situation, the application of AI for CXR has attracted more attention. AI-assisted diagnostic approach to the analysis of chest radiographs has been found to be beneficial and has the potential to be of assistance to physicians at point-of-care in resource-constrained countries.4-6 Concerning the reading time of radiologists, there could be a concern as to whether referring to AI results would increase workload by adding working steps or reduce decision-making time as an effective computer-assisted diagnosis system.7 In a prospective observational study7 (11 radiologists and 18,680 CXRs), how AI affects the actual reading times of radiologists in the daily interpretation of CXRs in real-world clinical practice has been evaluated. The authors7 reported that the availability of AI software influences the reading times of CXRs amongst radiologists and that AI integration can overall shorten reading times. However, it is important to note that abnormalities detected by AI may lengthen reading times since radiologists might take more time to judge the validity of the AI assessment and to report more details about the findings seen on images regardless of the accuracy of displayed AI results. However, the study reported several limitations and future directions. The study utilised only one commercially available software, and the generalisability of its results could be limited. Second, the number of CXRs containing lesions was different in the AI-unaided and aided periods. The authors7 did not perform the evaluation of whether the presence of lesions or the abnormality score was accurate according to the radiologists’ reports or computed tomography (CT) images. The study7 suggested investigating differences in reading times based on the experience and expertise of radiologists as an important area for future research. Another study3 provides a comprehensive and systematic overview of deep learning applications designed to facilitate CXR interpretation. The impact of AI-assisted interpretation on reading time has been evaluated, with evidence indicating that reporting time was variably affected.3 A demonstrable impact to patient outcomes may follow AI-enabled efficiency gains to radiology workflows; however, the study recommends further research to establish the presence or extent of such benefits. Furthermore, the authors discussed on the risks and safety highlighted by researchers, including the potential for poor model generalisability, suboptimal case labelling and the potential for data perturbation. Limitations in the availability of large, high-quality and accurately labelled CXR datasets can present a potential risk for developing and testing high-performing and appropriately generalisable ML models.8 Natural language processing (NLP) can be problematic and noisy when used for the generation of training or ground truth labels. For example, it has been reported that the NLP-generated labels in the CXR 14 dataset, which was used in 17 studies, do not accurately reflect the visual content of the CXR images.9 Apart from the risks involved, the benefits of AI include improved reporting accuracy which has the future potential to reduce false-positive and false-negative results and reduce unnecessary follow-up CT examinations. This may lead to early detection of findings and improved patient outcomes in screening, outpatient, emergency and inpatient settings. In a study,10 the device evaluated had focussed on detecting false negatives in CXRs originally interpreted as normal by radiologists using AI.10 In this study,10 the researchers had demonstrated a false referral rate of 0.97% and found that 1.2% contained salient clinical findings. Employing ML models to reduce false-negative rates and improve the quality of reporting in this way will continue to be of interest to radiology providers as workload volume and complexity grow. Generative AI methods have the ability to produce more informative and relevant outputs through the generation of the entire radiology report, providing important context for decision-making. However, clinically oriented evaluations of generative AI remain scarce in the biomedical literature. A study11 retrospectively evaluated a generative AI tool for chest radiograph report generation in the ER setting. The study concluded that AI-generated reports were not significantly different from radiologist reports, although they performed better than teleradiology reports, providing evidence for the applicability of AI to supplement ER physician decision-making in settings without immediate access to radiology services. However, there are some limitations that warrant consideration.11 In the real-life field situation in India, although the CXR radiograph facility is available to treating doctors at the primary healthcare level, these doctors do not have immediate access to expert radiologists’ opinions on the CXRs. In this regard, AI/ML methods can compensate for this deficiency to a large extent. Thus, the application of AI/ML methods for CXR reading can be handy as an expert advisor for physicians at primary healthcare services apart from minimising human error and fatigue. The scenario for application of AI/ML methods to other modalities such as CT and magnetic resonance imaging is different and merits further study.