Artificial intelligence–based systems demonstrated performance comparable to trained physicians in bronchoscopy, according to a systematic review of 20 studies. The findings suggested artificial intelligence's potential role in supporting clinical decision-making and training in pulmonary diagnostics.

The studies were grouped into three categories: airway anatomy and navigation (9 studies), computer-aided diagnosis (CADx) in endobronchial ultrasound (EBUS; 7 studies), and rapid on-site evaluation (ROSE; 4 studies). Investigators evaluated whether artificial intelligence (AI) could match or exceed human performance in identifying bronchial segments, classifying lymph nodes, or interpreting cytology samples.

Among the 16 assessment studies, 12 reported that AI performed comparably to human evaluators, whereas 4 demonstrated AI performing better in specific tasks. Four additional studies implemented AI in simulated settings, all within the airway navigation category. In each case, AI guidance improved procedural accuracy over unassisted performance.

In airway navigation, AI-based systems accurately identified bronchial segments using bronchoscopy images. One randomized trial found that novices using AI-guided training achieved inspection performance comparable to experienced bronchoscopists after several hours of practice. Another study showed that practitioners with over 500 procedures improved efficiency with AI support.

For EBUS applications, five studies found AI matched expert performance in classifying lymph nodes as benign or malignant. Two studies showed AI outperforming human evaluators. These systems were trained on large image data sets, some exceeding 2 million EBUS images. However, most results were based on static image evaluations rather than full procedural testing.

In ROSE, three studies found AI as accurate as expert cytopathologists in evaluating biopsy samples. One study reported inferior performance. Only one study included external validation, which revealed a modest performance decline compared with internal testing.

Study quality was assessed using a standardized scoring tool. The average methodologic quality was moderate, with randomized controlled trials, particularly those on AI-assisted training, receiving the highest scores.

Despite encouraging results, few studies tested AI in real clinical settings. Only four implementation studies were conducted, all within airway navigation. The investigators called for more clinical trials to determine AI’s effectiveness in improving patient outcomes during bronchoscopy.

The review found that AI systems often matched or exceeded expert performance in key bronchoscopy tasks. Further research is needed to validate these systems in clinical environments and ensure consistent reporting standards across studies.

The authors declared no conflict of interest.

Source: European Respiratory Review