甲状腺穿刺吸引生検の適格性評価のための拡張現実人工知能ツールの実現可能性

To evaluate the feasibility of developing an artificial intelligence application for real-time specimen adequacy assessment during ultrasound-guided fine needle aspiration of the thyroid gland using small, locally derived datasets and open-source AI tools.

1,244 photomicrographic images of thyroid needle aspiration slides were used to develop training and testing datasets for a YOLOv5 convolutional neural network model. In silico performance was tested for standard metrics. The model was deployed in an application featuring a live, augmented reality microscopy display. Hands-on testing was performed using a set of 50 blinded patient specimens tested in duplicate (N=100) by human observers with no prior formal training in pathology.

On an independent static image test set, the CNN achieved an AUROC of 0.954 (95% CI, 0.922-0.986) using images acquired with a dedicated microscope camera, with sensitivity 0.969, specificity 0.933, and accuracy 95.5%. Using images acquired with a smartphone camera, performance improved to an AUROC of 0.983 (95% CI, 0.965-1.000), with sensitivity 0.980, specificity 1.000, and accuracy 98.8%. In a simulated clinical workflow, human users assisted by the augmented reality AI tool demonstrated 96% concordance with the reference standard FNA adequacy assessment.

An augmented reality microscopy tool based on a small-model CNN showed feasibility for performing real-time assessment of thyroid FNA specimen adequacy. Such a tool can enable telecytologic support of FNA assessment in radiologic facilities without an on-site pathologist.