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

Area of Science:

• Oncology
• Medical Diagnostics
• Artificial Intelligence

Background:

• Fine needle aspiration (FNA) biopsy is crucial for thyroid nodule diagnosis.
• Approximately 10% of initial thyroid FNA samples are non-diagnostic, necessitating repeat procedures and delaying care.
• Current on-site evaluation of FNA samples requires time-consuming staining and expert cytopathologist presence.

Purpose of the Study:

• To develop and validate a deep learning model (FNA-Net) for in situ screening of unstained thyroid FNA samples.
• To reduce the non-diagnostic rate of thyroid FNA biopsies.
• To bypass the need for staining and on-site expert interpretation, thereby streamlining the diagnostic process.

Main Methods:

• Developed FNA-Net, an ensemble deep learning model combining a patch-based whole slide image classifier and Faster R-CNN.
• The model detects follicular clusters in unstained FNA slides.
• Samples with fewer than a predetermined threshold of follicular clusters are classified as non-diagnostic.

Main Results:

• FNA-Net achieved an F1 score of 0.81 and an AUC of 0.84 in detecting non-diagnostic slides (fewer than six follicular clusters) using bootstrapped sampling.
• The model demonstrated high precision in identifying follicular clusters.
• The developed AI model shows promise in accurately assessing sample adequacy.

Conclusions:

• FNA-Net offers a rapid, automated method for assessing the adequacy of thyroid FNA samples without staining.
• The AI model has the potential to significantly decrease the non-diagnostic rate of thyroid FNA biopsies.
• Implementing FNA-Net can reduce diagnostic costs and enhance the quality of patient care for thyroid nodules.