Artificial intelligence in thyroid nodules and cancer: clinical validation and real-world performance

Artificial intelligence in thyroid ultrasound: clinical applications and perspectives

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.

Beyond traditional biopsy: integrating liquid biopsy, molecular panels, and advanced imaging techniques in thyroid cancer management.

The incidence of thyroid cancer has increased dramatically over the past four decades, making it the most common endocrine malignancy worldwide. Although well-differentiated thyroid carcinomas generally exhibit excellent survival rates following standard surgical and radioactive iodine interventions, the clinical management of high-risk variants, recurrent disease, and indeterminate nodules remains a significant challenge. Historically, the evaluation of thyroid nodules has relied on B-mode ultrasonography and fine-needle aspiration biopsy (FNAB). However, spatial and temporal tumor heterogeneity, coupled with the high rate of indeterminate cytology (Bethesda Categories III and IV), often leads to unnecessary diagnostic thyroidectomies. In recent years, liquid biopsy has emerged as a promising diagnostic approach, offering a non-invasive, dynamic, and comprehensive method for disease monitoring. This comprehensive review evaluates the biological mechanisms and clinical utility of circulating tumor DNA (ctDNA), circulating tumor cells (CTCs), extracellular vesicles (EVs), and non-coding RNAs in thyroid oncology. Furthermore, it explores the clinical role of commercial molecular diagnostic panels (e.g., Afirma GSC, ThyroSeq v3) in guiding surgical decisions. This review highlights the diagnostic synergy achieved when these molecular biomarkers are integrated with advanced radiological modalities - including strain elastography, quantitative region-of-interest (ROI) analysis, ACR TI-RADS, and cross-sectional volumetric tracking - as well as nuclear medicine techniques like technetium-99m MIBI scintigraphy. Emerging applications of artificial intelligence and radiomics in image analysis, alongside the concept of minimal residual disease monitoring, are also discussed. Ultimately, we emphasize the need for a multidisciplinary team (MDT) approach in translating these multi-omics and multimodal imaging data into precision medicine, while acknowledging current limitations and charting future directions.

THE ROLE OF ARTIFICIAL INTELLIGENCE IN THE DIAGNOSIS AND TREATMENT OF SOLITARY LUNG LESIONS

Solitary lung lesions are being detected increasingly often in modern clinical practice due to the widespread use of chest computed tomography (CT), the development of screening programs based on low-dose computed tomography (LDCT), and advances in medical imaging technologies. Although the majority of solitary pulmonary nodules are benign, a certain proportion may represent an early manifestation of lung cancer. Therefore, the main clinical challenge is to achieve early detection of malignant lesions while avoiding unnecessary invasive investigations and treatment of benign nodules.In recent years, artificial intelligence, machine learning, deep learning, and radiomics have emerged as im­portant approaches for the assessment of solitary lung lesions. Artificial intelligence is being used for the automated detection of pulmonary nodules on CT scans, their segmentation, measurement of size and volume, assessment of growth dynamics, differentiation between benign and malignant lesions, and prediction of the probability of malig­nancy. A 2024 meta-analysis demonstrated that externally validated deep learning–based computer-aided diagnostic models increased sensitivity by 11.6% compared with physician assessment and by 14.5% compared with clinical risk models alone.Radiomics enables the extraction of a large number of quantitative features from CT images that are difficult to fully assess by visual examination alone. According to a 2023 meta-analysis, CT radiomics models demonstrated an overall AUC of 0.91, sensitivity of 0.86, and specificity of 0.84 for predicting the malignancy of pulmonary nodules. However, methodological limitations and a high risk of bias were reported in the majority of studies.The application of artificial intelligence is not limited to diagnosis and is expanding to the selection of follow-up intervals, automated detection of nodule growth, bronchoscopic navigation, surgical planning, and determination of treatment volumes in radiation therapy. At the same time, insufficient external validation of algorithms, differences between populations, variations in CT protocols, the “black box” problem, algorithmic errors, data security, and issues of clinical responsibility remain important barriers to the widespread implementation of these technologies in clinical practice.

Artificial intelligence in breast cancer imaging: a systematic review of detection, segmentation, explainability, and clinical translation

Artificial intelligence shows substantial potential to enhance breast cancer imaging, but broader clinical translation requires robust external and prospective validation, improved calibration, assessment of generalizability and bias, and integration into clinical workflows.

[Artificial intelligence-enabled advances in gastric cancer imaging: current status, challenges, and prospective strategies].

Clinical value of artificial intelligence ultrasound-assisted diagnostic system in the differential diagnosis of benign and malignant thyroid nodules

Accurately distinguishing the benign thyroid nodules (BTNs) and malignant thyroid nodules (MTNs) is crucial for treatment planning and prognosis. This study aims to explore the clinical value of artificial intelligence ultrasound-assisted diagnostic system (AI-UADS) in diagnosing the BTNs and MTNs. In this retrospective case-control study, a total of 161 TNs from 91 patients confirmed by surgical pathology were included. All nodules underwent ultrasound examinations and were evaluated by both senior ultrasonographers and AI-UADS, independently. Using surgical pathology as the gold standard, the diagnostic performance of the 2 methods was analyzed using receiver operating characteristic curves (including classify the nodules based on size, location, capsule relationships, and presence or absence of Hashimoto thyroiditis [HT]). The area under the curve, accuracy, sensitivity, specificity, positive predictive value and negative predictive value of AI-UADS in the differential diagnosis of BTNs and MTNs were 0.933, 93.2%, 92.4%, 94.2%, 95.5%, and 90.3%, respectively, all of which were higher than ultrasonographers (0.821, 82.0%, 81.5%, 82.6%, 86.2%, and 77.0%, respectively) (all P < .05). Specifically, the accuracy in diagnosing nodules with a maximum diameter ≤10 mm (94.0% vs 80.7%) and in the left lobe (92.5% vs 81.3%); the accuracy, sensitivity, specificity, positive predictive value and negative predictive value in diagnosing nodules with far from the capsule ([93.8% vs 79.5%], [92.5% vs 75.5%], [94.9% vs 83.1%], [94.2% vs 80.0%], [93.3% vs 79.0%]) and without HT ([94.0% vs 82.7%], [93.2% vs 82.4%], [94.9% vs 83.1%], [95.8% vs 85.9%], [91.8% vs 79.0%]) of AI-UADS were higher than ultrasonographers (all P < .05). The AI-UADS demonstrates good diagnostic performance in differentiating BTNs and MTNs, and can serve as an effective auxiliary tool for ultrasonographers. It shows particular advantages in diagnosing nodules with a maximum diameter ≤10 mm, located in the left lobe, far from the capsule, and in patients without HT.