Screening adequacy of unstained thyroid fine needle aspiration samples using a deep learning-based classifier

Abstract Fine needle aspiration (FNA) biopsy of thyroid nodules is a safe, cost-effective, and accurate diagnostic method for detecting cancer. However, about 10% initial FNA samples from patients are non-diagnostic require repeated FNA, which delays the diagnosis appropriate care. On-site evaluation sample can be performed to filter out samples. Unfortunately, it involves time-consuming staining process, cytopathologist has present at time FNA. To bypass process expert interpretation specimens clinics, we developed deep learning-based ensemble model termed FNA-Net that allows in situ screening adequacy unstained smeared on glass slide decrease rate combines two learning models, patch-based whole image classifier Faster R-CNN, detect follicular clusters with high precision. Then, classifies slides if total number detected less than predetermined threshold. With bootstrapped sampling, achieved 0.81 F1 score 0.84 AUC precision-recall curve whose six. We expect dramatically reduce cost associated improve quality patient