Cytological Classification Diagnosis for Thyroid Nodules via Multimodal Model Deep Learning

Abstract

The rising prevalence of thyroid nodules is straining limited cytopathology resources, resulting in excessive overdiagnosis and overtreatment with significant patient and healthcare consequences. To address this, AI-TFNA is developed, a robust artificial intelligence platform leveraging extensive clinical data to enhance diagnostic accuracy and clinical efficiency. A total of 20,803 thyroid samples are collected from seven medical centers across different regions in China. Of these, 4,421 thyroid fine-needle aspiration (TFNA) samples from three hospitals are used to train AI-TFNA, ensuring strong generalizability across diverse clinical settings. For the internal validation, AI-TFNA demonstrates exceptional performance: the overall accuracy of TBS I is 93.27%, the sensitivity of TBS V and TBS VI reaches 85.37% and 83.78%, while the specificity of TBS II is 97.13%. Consistent results are observed in an external cohort of 2,153 samples, demonstrating robust generalizability. The incorporation of BRAF mutation data into AI-TFNA and the development of a multi-modal model further improve precision by significantly improving the differentiation between benign and malignant thyroid nodules. Image Appearance Migration (IAM) is an innovative technique that substantially improves cross-institutional model generalizability, increasing AI-TFNA sensitivity by 1.90% and specificity by 8.12%. AI-TFNA offers rapid, reliable decision support, advancing thyroid nodule diagnostics.