Aug 2026· Frontiers in Endocrinology· Vol 17· 0 citations· 138 references
Medicine
TL;DR
This review summarizes AI applications in thyroid ultrasound, including image preprocessing, nodule segmentation, quantitative feature analysis, benign-malignant differentiation, TIRADS optimization and automated reporting, and highlights AI’s potential in enhancing diagnostic consistency and accuracy.
Abstract
Thyroid nodules are highly prevalent, with increasing detection rates driven by advanced imaging and expanded screening. Ultrasound serves as the first-line tool for screening, diagnosis and follow-up, owing to its non-invasiveness, real-time capability, cost-effectiveness and absence of ionizing radiation. However, conventional ultrasound diagnosis is highly operator-dependent, resulting in substantial inter-observer variability and diagnostic errors, particularly for subtle or indeterminate lesions. Artificial intelligence (AI), particularly deep learning and radiomics, has emerged as a promising approach to address these limitations by enabling automated feature extraction, quantitative analysis and standardized interpretation, which has the potential to improve diagnostic efficiency and risk stratification. This review summarizes AI applications in thyroid ultrasound, including image preprocessing, nodule segmentation, quantitative feature analysis, benign-malignant differentiation, TIRADS optimization and automated reporting. We highlight AI’s potential in enhancing diagnostic consistency and accuracy, while critically assessing the methodological quality, bias risks and external validation of existing studies. Most AI tools are still in early translational phases, lacking large-scale validation in real clinical settings and standardized reporting protocols. We further discuss key challenges, including data bias, limited generalizability due to small or single-center datasets, poor interpretability and significant translational barriers. Future directions involving multi-modal fusion, explainable AI, real-time clinical systems and rigorous, multi-center standardized validation are proposed to facilitate clinical translation and improve patient care.
Thyroid nodules are detected in a large proportion of adults undergoing high-resolution ultrasonography, yet only a minority harbor clinically significant cancer. The clinical problem is therefore not only cancer detection but calibrated risk stratification: avoiding delayed diagnosis of aggressive disease while limiting unnecessary biopsies, molecular testing and diagnostic surgery. Artificial intelligence (AI) has moved rapidly from experimental image classification to clinically-deployed decision support. This invited review synthesizes current evidence for AI applications in the evaluation and management of thyroid nodules and differentiated thyroid cancer, emphasizing ultrasound-based computer-aided diagnosis, indeterminate cytology, molecular integration, cytopathology and histopathology, lymph-node assessment, report quality control, surveillance and emerging multimodal large language models. Commercial and near-commercial systems including S-Detect, AmCAD-UT, Koios DS Thyroid, AIBx and newer deep-learning systems show that AI can improve consistency, support less experienced readers and, in selected settings, reduce low-yield fine-needle aspiration without unacceptable loss of sensitivity. A particularly important future role may be AI-enabled de-escalation, in which image-derived estimates of benignity help support surveillance when clinical, sonographic, cytologic or molecular risk signals are concordantly low. However, performance varies by case mix, cancer prevalence, scanner platform, operator experience, geographic cohort, reference standard and whether the model is used as a stand-alone classifier or second reader. The strongest evidence supports AI as an adjunct to standardized ultrasound risk stratification and shared decision-making, not as a replacement for expert clinical judgment. Future progress will depend on prospective multicenter validation, transparent reporting, local calibration, workflow design, regulation, post-market surveillance and assessment of patient-centered outcomes.
Mustafa Şahin, N. Angelopoulos, R. Paparodis· Endocrine Connections· 0 citations
A comprehensive narrative review of current AI/DL applications in ultrasound diagnosis, synthesising evidence across four major clinical domains and identifying recurring limitations across the literature - dataset heterogeneity, limited external/multicentre validation, and interpretability gaps - and outlines directions for future research toward clinically deployable, trustworthy AI-assisted ultrasound diagnosis.
R. Sivakumar, Arati Shahapurkar, J. G et al.· Adolescência e Saúde· 0 citations
The continued integration of AI has the potential to improve diagnostic accuracy, streamline radiological workflows, and support more personalized patient care in the next generation of diagnostic imaging.
Mr. Vishal Walia, Mr. Honey Thakur, Ms. Ashwarya Sharma et al.· PAIN, JOINTS, SPINE· 0 citations
Thyroid nodules are increasingly detected incidentally, with prevalence rates of 19%–68% in ultrasound studies, yet only 7%–15% harbor malignancy. Traditional risk stratification systems demonstrate significant inter-observer variability and modest diagnostic accuracy. Artificial intelligence (AI) and machine learning technologies have emerged as powerful tools for enhancing thyroid nodule evaluation and enabling more precise risk-based clinical management. Deep learning algorithms, particularly convolutional neural networks, demonstrate diagnostic accuracy of 83%–97% in differentiating benign from malignant thyroid nodules on ultrasound, often matching or exceeding expert radiologists. AI systems analyzing cytopathology images achieve a sensitivity of 87%–99% and specificity of 71%–97% in predicting malignancy from fine-needle aspiration specimens. Integration of radiomics features, molecular markers, and clinical data through machine learning models enables personalized risk prediction with area under the curve values exceeding 0.90. These technologies promise to reduce unnecessary biopsies and surgeries, minimize patient anxiety, and optimize resource utilization. However, implementation challenges include limited external validation, algorithmic transparency concerns, regulatory considerations, and the need for prospective clinical trials. This narrative review examines current AI applications in thyroid nodule risk stratification, analyzes performance metrics across different modalities, discusses clinical implications for early detection and optimized management, and explores future directions, including multimodal integration and real-time clinical decision support systems. As AI technology matures, its role in transforming thyroid nodule evaluation from population-based screening to personalized risk-stratified surveillance appears increasingly promising.
A. Almohammadi· Thyroid Research and Practic...· 0 citations
AI should be viewed not as a replacement for ultrasound professionals but as a decision-support technology that may improve consistency, efficiency, and diagnostic confidence when carefully validated in real-world clinical settings.
Bayan Alghamdi, Eman M Alrewily, Sharefa S Alghamdi· Ultrasound Quarterly· 0 citations
Breast cancer represents a leading public health problem due to its high incidence, morbidity, and mortality. Significant geographic differences in disease burden are influenced not only by socioeconomic and behavioural factors but also by the resources of health systems, access to healthcare, and the quality of diagnostic programs. Modern mammography screening programs substantially contribute to early disease detection, while existing challenges include limited resources, restricted access and inequalities in healthcare, high costs, variability in result interpretation, and the need for additional and invasive diagnostic procedures. To address these challenges, artificial intelligence systems have been developed, employing complex algorithms, machine learning, and deep learning for the automated analysis of mammographic images. These systems can improve diagnostic accuracy, reduce reading time, lower false-negative rates, and decrease the need for repeat diagnostic procedures. Research has demonstrated that the accuracy achieved through artificial intelligence algorithms is comparable to that of radiologists, and optimal results may be obtained by combining radiologist's assessments with artificial intelligence methods, particularly in the role of the first reader. Key limitations of these methods include insufficient adaptability to diverse populations and the risk of over-reliance on algorithm results by less experienced radiologists. Beyond technical challenges, an important public health aspect concerns women's attitudes toward the use of artificial intelligence, as the level of trust in this technology may affect screening participation rates. The integration of artificial intelligence into screening programs requires careful evaluation of benefits and limitations, transparency in technology use, and assurance of human oversight. Proper implementation and public education can contribute to improved early diagnosis, greater screening accuracy, and more efficient use of health system resources.
Kristina Stamenković, V. Mijatović-Jovanović, D. Milijašević· Glasnik javnog zdravlja· 0 citations