Machine Learning Predicts Adequacy of Rapid Onsite Evaluation in Fine Needle Aspirations in Lung Cancer Cytology

Lung cancer is projected to become the leading cause of cancer-related mortality in both smoking and non-smoking populations. Rapid onsite evaluation (ROSE) of fine-needle aspiration specimens is essential for timely diagnosis and procedural decision-making during lung cancer assessment. We developed a machine-learning pipeline for cell-based adequacy assessment and lesion detection that integrates automated cell detection, convolutional neural network-based cell classification, and slide-level aggregation using a random forest model. On held-out test data, binary classifiers for lymphocytes and tumor cells achieved accuracies of 91.5% and 92.7% with recalls of 92.6% and 93.1%, respectively. The end-to-end ROSE system demonstrated class accuracies of 82-85%, comparable to human cytologist performance, and a lesion-focused classifier reached a recall of 92.0%. These findings indicate that machine-learning-based cell analysis can support ROSE by expediting adequacy assessment and improving diagnostic yield during TBNA procedures.