Rapid On-Site Evaluation is a key quality tool in interventional pulmonology. ROSE allows immediate cytologic assessment of biopsy specimens during procedures such as endobronchial ultrasound-guided transbronchial needle aspiration and transthoracic needle aspiration. This real-time adequacy check reduces the need for repeat biopsies and improves procedural efficiency and patient experience.
Most centers lack dedicated on-site cytopathologists. Not all specialty centers can guarantee the presence of an experienced cytologist during interventional pulmonology sessions. When pathologists are unavailable, centers rely on trained technicians or pulmonologists, but published data show sensitivity for ROSE performed by pulmonologists ranges from 65% to 91%, indicating meaningful variability in quality.
Telemedicine and AI offer potential solutions to the pathologist shortage. Remote consultation via encrypted digital platforms enables pathologists to evaluate biopsy images in real time without being physically present. Simultaneously, machine learning algorithms trained on cytology images could assist pulmonologists in identifying pathological cells and reducing inter-operator variability.
This study directly compared E-ROSE and AI-ROSE against the gold standard. The multicentric study evaluated 277 biopsy slide images from multiple sampling methods, comparing remote pathologist evaluation via encrypted messaging (E-ROSE) and machine learning classification (AI-ROSE) against the definitive histological diagnosis supplemented by immunohistochemistry.
Biopsy images were collected from multiple sampling techniques across two centers. Images were obtained by EBUS-guided transbronchial needle aspiration, endoscopic ultrasound-guided fine needle aspiration, echo-guided transthoracic needle biopsy, and transbronchial biopsy. Slides were stained with the May-Grunwald Giemsa method and imaged at 10x and 20x magnifications on a Leica microscope.
E-ROSE used encrypted image sharing with an expert remote pathologist. An experienced pulmonologist selected one to five representative images per case and transmitted them via encrypted platforms to a remote anatomic pathologist with over 10 years of lung cytopathology experience. The pathologist classified each sample as positive, doubtful, or negative in real time using morphological criteria from the Pulmonary Pathology Society guidelines.
AI-ROSE used Google Inception v3 and multiple machine learning classifiers. Images were processed through the Inception v3 model pre-trained on ImageNet to extract 2,048 features per image. Fast Correlation-Based Filter dimensionality reduction then identified 44 relevant features for the positive/negative analysis and 55 for the analysis including doubtful cases. Algorithms tested included neural networks, naive Bayes, logistic regression, k-nearest neighbors, CN2 rule induction, and stochastic gradient descent, each validated by 10-fold cross-validation.
Two analyses were conducted: with and without doubtful cases. The first analysis classified 252 slides as strictly positive or negative. The second added 25 doubtful cases for a total of 277 slides, allowing assessment of model robustness under real-world ambiguity. The final gold standard was the definitive histological report including immunohistochemistry, against which all E-ROSE and AI-ROSE results were compared.
E-ROSE achieved near-expert diagnostic accuracy when excluding doubtful cases. For the 252 definitively classified slides, E-ROSE demonstrated sensitivity of 99.0%, specificity of 88.7%, accuracy of 95.5%, and an AUC of 93.9%. The F1 score was 97.1% and the Matthew correlation coefficient was 90.7%, indicating excellent classification reliability by the remote expert pathologist.
Including doubtful cases reduced but maintained strong E-ROSE performance. When the 25 ambiguous slides were added, E-ROSE sensitivity remained high at 97.1%, while specificity fell to 81.0% and overall accuracy to 91.4%. AUC decreased to 89.1%, reflecting the natural difficulty of evaluating morphologically ambiguous specimens through static images without direct microscopic control.
AI-ROSE neural networks performed best for clear positive versus negative classification. The neural network model with three hidden layers achieved sensitivity of 96.4%, specificity of 78.9%, accuracy of 92.5%, and AUC of 94.8% for the positive/negative analysis. The F1 score of 92.3% and Matthew correlation coefficient of 77.9% confirmed strong but slightly lower performance than E-ROSE.
The k-nearest neighbors algorithm outperformed the neural network when doubtful cases were included. When the full 277-slide dataset including doubtful cases was analyzed, the kNN model with k equals 5 and Euclidean distance achieved sensitivity of 97.4%, specificity of 75.4%, accuracy of 85.2%, and AUC of 82.3%. Both E-ROSE and AI-ROSE AUC values converged to similar levels when doubtful cases were included (89.1% vs 86.4%), suggesting comparable real-world discriminative ability.
E-ROSE addresses the fundamental problem of pathologist availability. The study demonstrates that a remote pathologist can provide cytological evaluation comparable in accuracy to physical on-site presence. Since the pathologist's direct involvement is needed only during specific, brief phases of the procedure, remote digital consultation represents a more efficient use of specialist resources than requiring full on-site presence for the entire procedure duration.
AI-ROSE provided image classification in approximately 30 seconds. Based on clinical experience with the system, AI-ROSE delivered results in roughly 30 seconds while E-ROSE required 1 to 3 minutes depending on image transmission time and pathologist availability. This speed advantage may be clinically relevant in time-pressured settings and for rapid assessment of sample adequacy during ongoing procedures.
Both methods compare favorably to existing published benchmarks. Prior published AI studies on EBUS-TBNA ROSE showed sensitivity of 95.1% and specificity of 93.3% in the DEBUT study, and accuracy of 87% to 90% in the Yan et al. study. The current results fall within this range for AI-ROSE, while E-ROSE exceeded most published AI benchmarks. Importantly, in prior studies as in this one, the expert pathologist outperformed AI, but AI outperformed junior and intermediate-level endoscopists.
The economic case for E-ROSE and AI-ROSE is favorable. Direct material costs per procedure were approximately equivalent between standard ROSE and AI-ROSE at approximately $888 per case. Published cost analyses indicate that telepathology systems become economically advantageous when monthly case volumes exceed certain thresholds, suggesting these approaches are best suited to moderate to high-volume interventional pulmonology units.
Pathologists can continue laboratory work while providing remote ROSE support. The E-ROSE model enables an expert pathologist to maintain their routine laboratory activities while simultaneously consulting on biopsy adequacy from a remote location. This dual-use capability makes specialist expertise available to more centers without requiring additional full-time positions or physical relocation of staff.
Digital image capture enables training and quality improvement. The ability to store and share diagnostic images on digital platforms creates a growing repository of labeled cytological specimens that supports specialist training, quality control, and refinement of AI models. Trainees can access high-quality educational cases through mobile devices, reducing the resource intensity of cytopathology training programs.
Combining E-ROSE and AI-ROSE could represent an optimal diagnostic strategy. The authors propose that the two approaches are complementary rather than competing. AI-ROSE can provide an immediate initial classification, flagging cases as clearly positive, clearly negative, or potentially ambiguous for expedited human review. E-ROSE then provides expert human oversight for doubtful cases, combining speed and accuracy in a clinically practical workflow.
Broader adoption requires validation on larger and more diverse datasets. The primary limitation of AI-ROSE is the relatively small number of training and validation images in this study. Machine learning models for cytological interpretation need extensive diverse datasets to handle the full range of morphological variation encountered in practice, including samples with low cellularity, atypical morphology, or technical quality issues such as poor focus or non-uniform staining.
E-ROSE maintains diagnostic quality without requiring on-site pathologist presence. With overall accuracy of 91.4% including doubtful cases and 95.5% for clear cases, E-ROSE demonstrated that remote encrypted image sharing is a reliable alternative to on-site pathology review. This approach could fundamentally change how interventional pulmonology units are staffed and organized.
AI-ROSE reduces inter-operator variability and standardizes diagnostic criteria. By applying consistent morphological classification rules derived from training data, AI algorithms reduce the subjectivity that currently causes variation in ROSE quality across centers and operators. This standardization benefit extends beyond diagnostic accuracy to the reproducibility of adequacy assessments across different operators and institutions.
Implementation requires attention to image quality and case complexity. Limitations of both approaches include their dependence on image quality, potential sampling bias from static image selection by pulmonologists, and reduced effectiveness for cytologically ambiguous samples. These factors are most pronounced in low-cellularity specimens and cases with atypical morphological features that challenge even expert pathologists.
Prospective studies in diverse settings are needed to confirm generalizability. This multicentric retrospective study provides strong proof-of-concept evidence for both E-ROSE and AI-ROSE, but prospective validation across a broader range of centers including community hospitals and low-resource settings is required before these methods can be recommended as standard of care for lung cancer cytological evaluation.