Annotation-free genetic mutation estimation of thyroid cancer using cytological slides from multi-centers

Abstract

Abstract Thyroid cancer is the most common form of endocrine malignancy and fine needle aspiration (FNA) cytology is a reliable method for clinical diagnosis. Identification of genetic mutation status has been proved efficient for accurate diagnosis and prognostic risk stratification. In this study, a dataset with thyroid cytological images of 310 indeterminate (TBS3 or 4) and 392 PTC (TBS5 or 6) was collected. We introduced a multimodal cascaded network framework to estimate BARF V600E and RAS mutations directly from thyroid cytological slides. The area under the curve in the external testing set achieved 0.902 ± 0.063 and 0.801 ± 0.137 AUCs for BRAF, and RAS, respectively. The results demonstrated that deep neural networks have the potential in cytologically predicting valuable diagnosis and comprehensive genetic status.

Keywords

• Deep convolutional network
• Thyroid cancer
• Fine needle aspiration cytology
• Annotation-free
• Gene mutation estimation