Abstract Rationale Bronchoscopic biopsies are essential for the diagnosis of lung cancer. Optimizing tissue acquisition for both diagnostic accuracy and adequate volume for ancillary molecular testing remains critical. Although rapid on-site evaluation can improve diagnostic yield, its reliability depends on cytopathology expertise and availability. Dynamic cell imaging (DCI) is a novel microscopy platform that detects intracellular motion as a surrogate for metabolic activity. This technology provides rapid imaging feedback of small samples that are intended to be read by the proceduralist in real-time during biopsy procedures, negating the need for a cytopathologist or technician. This may enable proceduralists to better identify malignant tissue in real time themselves, but its diagnostic performance in this setting has not been evaluated. Methods Using the first-generation Van Gogh System (Celltivity Scientific, Waltham, MA), DCI images of mediastinal lymph node aspirates and peripheral lung nodule biopsies were obtained from consecutive patients undergoing bronchoscopy. An experienced lung pathologist first reviewed all DCI images to establish reference diagnoses (presence or absence of malignancy) using strict cytologic criteria. The same DCI images were then shown in random order to interventional pulmonologists, who independently classified each sample as malignant or not, and rated their diagnostic confidence (0-100). Each reader was provided limited training prior to evaluation. Pooled concordance between pulmonologist and pathologist was calculated. A mixed-effects logistic regression model, accounting for the crossed random effects of reader and case, was used to estimate diagnostic performance. Model-based predicted probabilities of reader’s decision of malignancy were used to generate receiver operating characteristic (ROC) curves. Results Samples from 39 consecutive bronchoscopic cases were independently read by 9 interventional pulmonologists. Pooled concordance between pulmonologist and pathologist was 67.9%. After adjusting for reader confidence, case order, and the crossed random effects for reader and case, the odds of correctly identifying a truly malignant specimen were 42.5 times higher than for non-malignant cases (95% CI 1.88 - 963.47). The area under the model-based ROC curve was 0.880 (95% CI, 0.847-0.913), reflecting excellent overall discrimination (Figure 1). Conclusions DCI enabled interventional pulmonologists to identify malignancy with fair overall accuracy relative to the reference diagnosis. While model-based estimates demonstrated strong discriminative performance, some cases remained inherently difficult to interpret. This difficulty in discrimination could potentially be overcome by additional sampling and imaging at the same location. Larger prospective studies are warranted to validate DCI as a reliable adjunct for specimen assessment during bronchoscopic biopsies. This abstract is funded by: None
Latifi et al. (Fri,) studied this question.