ARTIFICIAL INTELLIGENCE IN ULTRASONOGRAPHIC DIAGNOSIS OF THYROID NODULES: ENHANCING RISK STRATIFICATION AND CLINICAL DECISON-MAKING

Abstract

Thyroid nodules are common clinical findings, increasingly detected due to the widespread use of high-resolution ultrasound imaging. While the majority of these nodules are benign, a minority may be malignant, necessitating accurate and efficient risk stratification. Traditional ultrasonographic evaluation relies heavily on the operator’s expertise and subjective interpretation, which introduces diagnostic variability. This narrative review explores the evolving role of artificial intelligence in the ultrasonographic diagnosis of thyroid nodules. The principal objective of this review is to critically evaluate the diagnostic performance, clinical utility, and integration potential of artificial intelligence (AI)-based methodologies—including machine learning (ML) and deep learning (DL)—in the ultrasonographic assessment of thyroid nodules. Particular attention is devoted to the enhancement of existing risk stratification frameworks, and to identifying barriers to implementation in routine clinical. The review evaluates AI-integrated diagnostic systems in relation to existing classification frameworks, such as the thyroid imaging reporting and data system, and highlights innovations in elastography, 3D imaging, and automated segmentation. Evidence suggests that AI can enhance diagnostic accuracy, reduce interobserver variability, and improve the standardization of thyroid nodule assessment. Some algorithms demonstrate performance comparable to that of experienced clinicians, particularly in differentiating benign from suspicious nodules. Despite promising results, limitations such as model generalizability, the need for large annotated datasets, and clinical validation remain challenges. The findings support the integration of artificial intelligence as a complementary tool to assist healthcare professionals in making more objective, consistent, and timely decisions regarding the evaluation and management of thyroid nodules Methodology: A comprehensive narrative review of peer-reviewed literature was undertaken, encompassing both classical and AI-augmented ultrasonographic techniques, with a specific focus on diagnostic criteria, algorithmic accuracy, and classification consistency across TI-RADS variants (ACR-TIRADS, EU-TIRADS, K-TIRADS). Additionally, the role of emerging modalities such as ultrasound elastography was examined in the context of evaluating cytologically indeterminate nodules. Literature published between 2009 and 2025 was examined to assess how machine learning and deep learning algorithms contribute to image interpretation, classification, and malignancy prediction. Abbreviated Description of The State Of Knowledge: Thyroid nodules are detected in up to 60% of the general adult population via ultrasonography. Although the malignancy rate remains relatively low (~5%), the clinical imperative is the accurate differentiation of malignant from benign lesions. Risk stratification relies on the assessment of sonographic features, including echogenicity, shape, margins, calcifications, and vascularity. Several standardized scoring systems—most notably TI-RADS—are employed to systematize malignancy risk and guide indications for fine-needle aspiration biopsy (FNAB). Despite its utility, ultrasonography remains inherently operator-dependent and subject to interpretive variability. AI-powered diagnostic systems have demonstrated promising potential in mitigating interobserver discrepancies, augmenting risk classification fidelity, and improving diagnostic throughput. Adjunctive techniques such as elastography provide additional biomechanical data, although limitations in methodological standardization currently preclude widespread adoption.

Ähnliche Arbeiten