CytoThyro AI: automated classification of thyroid FNAC images using deep learning based on 2023 Bethesda System for reporting thyroid cytopathology

Thyroid fine-needle aspiration cytology (FNAC) is the standard initial investigation for thyroid nodules, reported using the six-category 2023 Bethesda System for Reporting Thyroid Cytopathology (TBSRTC). Inter-observer variability, subjectivity, and cytopathologist shortages limit reporting consistency, creating a need for AI-assisted classification tools. In this prospective study, 175 patients from eastern India underwent thyroid FNAC. Digital microphotographs in May-Grünwald Giemsa (MGG) and Papanicolaou (Pap) stains at 10X and 40X magnifications were acquired, generating a total of 3,500 images. Bethesda category labels (B1–B6) were assigned by majority consensus of three blinded cytopathologists, forming a cytological rather than histopathological reference standard. Five CNN architectures were trained using patient-level data partitioning (70:15:15). Final models were selected using TBSRTC 2023 clinical safety criteria and combined via a weighted soft-voting ensemble. On a held-out test set of 27 patients, the Pap-based ensemble achieved patient-level classification accuracy of 96% (26/27; 95% CI 0.82–0.99; macro F1: 0.94; weighted F1: 0.96), with recall of 1.00 for B1, B2, B4, B5 and B6, and 0.50 for B3. Category-specific recalls of 1.00 rest on two to thirteen patients and carry wide confidence intervals. The MGG-based ensemble achieved 89% accuracy (24/27; 95% CI 0.72–0.96; macro F1: 0.85; weighted F1: 0.88), with recall of 1.00 for B1–B4, 0.33 for B5 and 0.80 for B6. At image level, AUC values across the four selected models ranged from 1.00 for B1, 0.95–0.99 for B2, 0.81–1.00 for B4, and 0.91–1.00 for B6. B3 (Atypia of Undetermined Significance) was the most challenging category, with image-level AUC ranging from 0.64 to 0.97 and image-level F1 from 0.31 to 0.71. Neither pipeline produced B1/B2-to-B5/B6 misclassification at patient level. The ensemble was deployed as CytoThyro AI, a publicly accessible web-based tool (cytothyro.cloud), intended at this stage for research and educational exploration.