A study on ultrasound imaging for thyroid detection and classification using machine learning and deep learning techniques

Abstract

Thyroid diseases are increasingly prevalent and have a major influence on endocrine and metabolic functions. Ultrasound imaging is popular for early diagnosis and categorizing thyroid abnormalities because it is non-invasive, inexpensive, and real-time. From this paper, we planned to study the recent developments in computer-aided diagnosis (CAD) systems in detecting and classifying thyroid diseases based on ultrasound images. Researchers are focused on traditional image processing algorithms, machine learning techniques, and deep learning techniques to improve accuracy. Modern classifiers like convolutional neural networks (CNNs) have garnered much attention in nodule segmentation, benign/malignant classification, and feature extraction. Our study encompasses major methodologies, datasets, performance measures, and challenges. Our investigation seeks to integrate a comprehensive view of existing studies in this area, point out existing gaps, and make recommendations on future research in thyroid ultrasound image analysis. From our survey, we concluded that the Deep learning models have shown superior performance relative to traditional techniques in several benchmark studies with very good accuracy, sensitivity, and specificity.

Resumen

Las enfermedades tiroideas son cada vez más prevalentes y tienen una gran influencia en las funciones endocrinas y metabólicas. La ecografía es popular para el diagnóstico temprano y la categorización de anomalías tiroideas porque es no invasiva, económica y en tiempo real. En este trabajo se estudian los avances recientes en los sistemas de diagnóstico asistido por computadora (CAD) para la detección y clasificación de enfermedades tiroideas a partir de imágenes ecográficas. Los investigadores se han centrado en algoritmos tradicionales de procesamiento de imágenes, técnicas de aprendizaje automático y técnicas de aprendizaje profundo para mejorar la precisión. Clasificadores modernos como las redes neuronales convolucionales (CNN) han recibido gran atención en la segmentación de nódulos, la clasificación benigno/maligno y la extracción de características. Nuestro estudio abarca las principales metodologías, conjuntos de datos, medidas de desempeño y desafíos. La investigación busca integrar una visión integral de los estudios existentes en esta área, señalar las brechas actuales y hacer recomendaciones para futuras investigaciones en el análisis de imágenes ecográficas tiroideas. De nuestra revisión concluimos que los modelos de aprendizaje profundo han mostrado un desempeño superior en comparación con las técnicas tradicionales en varios estudios de referencia, con muy buena precisión, sensibilidad y especificidad.

Introduction

Thyroid disease is a prevalent endocrine disease affecting thyroid gland function (a small, butterfly-like gland) at the front of the neck. The thyroid is responsible for controlling metabolism through the production of thyroid hormones. There are several types of thyroid disorders, each impacting the functioning of the thyroid differently. The most prevalent are hypothyroidism, in which too little hormone is secreted by the thyroid, and hyperthyroidism, in which too much is. Goiter is a swelling of the thyroid gland and is associated with both hypo- and hyperthyroidism. Thyroid nodules are nodules or tumors in the thyroid gland, either benign or cancerous. Though less prevalent, thyroid cancer is a severe disorder requiring early diagnosis and treatment. All thyroid disorders have different signs and symptoms, causes, and treatments, and therefore, diagnostic accuracy is crucial in their management. Hypothyroidism, hyperthyroidism, goiter, and thyroid cancer will occur if the gland creates too little or too much hormone or has unusual lumps called nodules.

There are different tests to detect thyroid abnormalities. Blood tests, like a test to measure levels of TSH, T3, and T4, are employed to check thyroid function. Radioactive iodine uptake tests to measure how much iodine is absorbed by the thyroid help check the activity of the gland. Fine-needle aspiration biopsy is normally carried out to take tissue samples of thyroid nodules to check cytologically. Imaging studies are equally important to check the anatomy, volume, and nature of the thyroid. These diagnostic tools combine to offer a clear picture of thyroid health. For example, imaging studies such as ultrasound can be used to detect nodules, whether they are solid or fluid-filled, and to guide needle biopsy in a precise fashion. When cancer of the thyroid is suspected, additional imaging such as CT or MRI is used to evaluate the extent of spread of the disease. The types of imaging tests available for diagnosing thyroid abnormalities are listed below:

• 1.

  Ultrasound¹ – High-frequency sound waves are used to create real-time images of the thyroid. A probe is moved over the neck area with gel applied to the skin.

• 2.

  Radioactive iodine uptake (RAIU)² – Radioactive iodine is swallowed or injected, and a special camera tracks how much iodine the thyroid absorbs over time.

• 3.

  Thyroid scan (scintigraphy) – Similar to RAIU, but includes imaging with a gamma camera to see where iodine is concentrated in the gland.

• 4.

  CT scan³ – Takes multiple X-ray images to produce detailed cross-sections of the neck and surrounding areas. Often used with contrast dye.

• 5.

  MRI⁴ – Uses strong magnets and radio waves to create detailed images without radiation. Often used if CT is inconclusive.

• 6.

  PET scan⁵ (occasionally used) – Injects radioactive glucose to detect areas of high metabolic activity, often used in cancer staging or recurrence detection.

Thyroid disease is one of the most widespread endocrine diseases across the globe, and early detection is crucial for preventing complications such as metabolic disorder as well as cardiovascular disease. Though ultrasound imaging remains the first preferred choice of diagnostic method for thyroid nodules due to its safety profile, ease of accessibility, and economic viability, its interpretation remains extremely operator-dependent and inconsistent. In the last few decades, artificial intelligence (AI) has emerged as a paradigm shift of image analysis that allows for automated feature detection, grading, and decision-making. When integrated into thyroid disease, AI approaches – particularly machine learning and deep learning – have been seen to be beneficial for improving diagnostic productivity with higher reliability, reducing inter-observer variation with consistent decision-making across several clinical scenarios. Such rising promise mandates that we present a discussion of AI-based approaches of thyroid disease identification and typing with regard to consideration of methodological patterns, data sets, performance, and limitations.

Correct diagnosis is important not only to determine the type of thyroid disease but also to individualize successful treatments. Based on the condition, treatments might be hormone replacement therapy, anti-thyroid medication, radioactive iodine therapy, or even removal of part or the entire thyroid. Early diagnosis and management can greatly improve outcomes and quality of life in people with thyroid disease.

Role of machine learning algorithms for analyzing thyroid abnormalities

Machine learning (ML) algorithms have a revolutionary role to play in characterizing thyroid abnormalities, especially through the classification of benign and malignant ultrasound-based nodules. The early ML algorithms, such as support vector machines (SVM), k-nearest neighbors (k-NN), and Decision Trees, were successful when combined with manually crafted features such as texture descriptors (e.g., GLCM, wavelets, histograms), obtaining satisfactory diagnostic performance. The introduction of deep learning – specifically Convolutional Neural Networks (CNNs) – has massively improved accuracy, sensitivity, and specificity in numerous investigations. By learning hierarchical image features end-to-end from data, these models surpass manual schemes and allow for high-precision classification, often obtaining more than 95% accuracy rates. Furthermore, hybrid and ensemble schemes combining CNNs with traditional classifiers (e.g., Random Forest, SVM) or adding attention mechanisms significantly improve model robustness and interpretability. While there are issues such as dataset variability and explainability of models, ML-based systems provide scalable, non-invasive, and reproducible diagnostic tools that assist clinicians in early and accurate detection of thyroid disease. Some of the effective machines learning algorithms for analyzing thyroid abnormalities from ultrasound images are listed below:

• 1.

  Supervised learning model (SVM) – It finds the optimal hyperplane for classification. This model used for Nodule classification (benign vs malignant), feature-based diagnosis.

• 2.

  Random Forest (RF) – An ensemble of Decision Trees used for classification and regression. Commonly used for Feature selection and thyroid disease classification.

• 3.

  K-nearest neighbors (k-NN) – Classifies a sample based on the majority label among its k-nearest neighbors. Commonly used for basic classification tasks (e.g., thyroid nodule labeling).

• 4.

  Naïve Bayes – Probabilistic classifier based on Bayes’ theorem with strong independence assumptions. It is used for early-stage classification using textural features.

• 5.

  Artificial neural networks (ANN) – Network of interconnected nodes mimicking brain neurons. Used for early detection and basic pattern recognition.

• 6.

  Convolutional neural networks (CNN) – A deep learning model is ideal for image data. Commonly used for nodule segmentation, feature extraction, benign/malignant classification.

• 7.

  Gradient boosting machines (GBM) – An ensemble technique that builds trees sequentially to reduce errors. Used for classification tasks in structured and semi-structured data.

While some recent reviews published syntheses on the application of machine learning in diagnosing thyroid diseases, these overwhelmingly tend to narrowly address either traditional classifiers or deep learning frameworks separately. A contrasting synoptic synthesis tracing historically the evolution of techniques from early hand-crafted feature-based to hybrid or ensemble or transformer-driven configurations along with identification of respective strengths, weaknesses, and clinical implications is provided by this synthesis. Balancing developments both in imaging modalities as well as in algorithmic developments, it distinguished itself in emphasizing not only diagnostic accuracy but also model robustness, explainability, and clinical translatability. Accordingly, it makes a distinguished contribution by bridging technological developments and practical implications in thyroid disease diagnosis and therapy.

Literature review

Data disclosure

For thyroid disease detection based on DL models, dataset quality and availability are also among the dominating factors that influence model construction and model verification. A well-compiled dataset not only improves model learnability of task-relevant features but also generalization towards unknown and new data. Of particular note, for functions like medical image-based thyroid nodule classification, such dataset details as the number of images, quality of annotation, and condition of scans add significantly towards achieving validity estimation and reproducibility of results studied. Highly diverse and accurate data also helps towards models learning the variability of nodules due to the device used during the scan and the patient's age, gender. Additionally, datasets are usually divided into training sets, verification sets, and test sets for model construction and model verification efficiently. This table is a compendium of dataset details available in various research articles used towards thyroid disease detection. This includes details related to sources of information, the number of total images available, details of images specifically, and types of ultrasound devices used in image capture. This overview, at a glance, presents informative details related to resources used in articles and promotes dataset transparency towards further investigations related to thyroid imaging. Table 1 gives the cumulative dataset used across various studies.

Table 1.

Dataset detail.

Author	Data source	No. of images	Image detail	Equipment used
D. Selvathi and V.S. Sharnitha (2011)^6	Partially obtained from online websites, partially collected from radiologists	50	30 images with thyroid nodules, 20 normal thyroid images	–
Acharya et al. (2013)^7	Not specifically named	200	100 normal images, 100 Hashimoto's thyroiditis	Logiq 5 (General Electric), 10 MHz linear transducer, harmonic imaging off
U. Rajendra Acharya et al. (2014)^8	Patients were from a Polish population.	526	232 normal images, 294 Hashimoto thyroiditis-affected images.	LOGIQ 5 (GE Healthcare)
Zbigniew Omiotek et al. (2015)^9	In-house dataset from 66 patients	139	54 samples from healthy cases and 85 from Hashimoto's cases	–
Ali Abbasian Ardakani et al. (2015)^10	Imam Khomeini Hospital, Urmia, Iran	70	26 benign, 44 malignant	Accuvix V20 (Medison, Korea); L5-13IS linear transducer (5–13 MHz)
Deepika Koundal et al. (2016)^11	Department of Radiology, Post Graduate Institute of Medical Education & Research (PGIMER), Chandigarh, India	42	16 males and 26 females. Age 15 to 70 years.	IU22 Philips X Matrix; linear probe, 17.5 MHz
Hanung Adi Nugroho et al. (2017)^12	RSUP Sardjito Hospital, Yogyakarta	39	25 cystic cases, 14 solid cases	–
Jianfu Xia et al. (2017)^13	Wenzhou Central Hospital, China	203 nodules	114 benign in 106 patients, 89 malignant in 81 patients	Acuson Sequoia 512 and 128XP (Siemens), 10–12 MHz linear probes
Jianguo Sun et al. (2018)^14	First Affiliated Hospital of Harbin Medical University	173	80 benign, 93 malignant	–
Eka Legya Frannita et al. (2018)^15	RSUP Sardjito Hospital, Yogyakarta, Indonesia	97	34 solid, 32 cystic, 31 complex images	–
Xueya Cen et al. (2019)^16	Hepatic hemangioma patients	3000	1200 sets of balanced positive and negative samples. 1000 images consisting of benign and malignant samples	–
Tianjiao Liu et al. (2019)^17	Cancer Hospital, Chinese Academy of Medical Sciences	8140	Dataset I: 7690 images from 4279 patientsDataset II: 450 images from 376 patients	GE Logiq E9, S7 ultrasound machines with 5–12 MHz or 8–15 MHz probes
Hervé Monpeyssen et al. (2020)^18	4 European thyroid referral centers. American Hospital of Paris (France), General Hospital of Livorno (Italy), Regina Apostolorum Hospital, Albano (Italy), Azienda Sanitaria di Teramo Hospital (Italy)	65	65 patients with 65 nodules	EsaoteMyLab Twice, LA435 for US; Echopulse system (Theraclion SA, France) for HIFU treatment
Mengya Liu et al. (2020)^19	Private dataset (not publicly named)	816	544 for training, 136 for validation	–
Ilah Shin et al. (2020)^20	Severance Hospital & Samsung Medical Center (South Korea)	348	252 nodules were confirmed as follicular adenoma.96 nodules were confirmed as follicular carcinoma.	–
Vadhiraj VV et al. (2021)^21	Open-source digital database shared by the Universidad Nacional de Colombia	134	99 patients and 134 ultrasound examinations, classified as 33 benign and 66 malignant cases based on TI-RADS scores.	–
Qingsong Wang et al. (2021)^22	Tianjin Medical University General Hospital	2032	Training set: 427 benign, 573 malignant. Test set: 100 benign, 100 malignant	–
Corina Maria Vasile et al. (2021)^23	Custom-collected ultrasound image dataset across four tertiary institutions in Craiova, Romania.	920	Autoimmune (260 images), Micro-nodular (215 images),Nodular (242 images), Normal (203 images)	Siemens SonoAce with a 7.5 MHz linear transducer
S. Pavithra et al. (2022)^24	Digital Database for Thyroid Ultrasound Images (DDTI)	328	Classified into 6 TI-RADS categories: TI-RADS 2 to TI-RADS 6	–
Rajshree Srivastava and Pardeep Kumar (2022)^25	Thyroid Digital Image Database (TDID), Kriti Scanning Center, Prayagraj, India, approved by NABH.	295, 654	TDID – 107 benign, 188 malignant. Scanning center – 428 benign, 226 malignant	Ultrasound machines: Voluson E-10, Mindray Resona 7, Siemens Healthineers, and Voluson E-8.Transducers: 3–11 MHz convex and linear transducers.
Shengsheng Yao et al. (2023)^26	China Medical University (Liaoning Provincial People's Hospital), Shenyang, China	106,513	Images from 17,934 patients showing various stages of Hashimoto's thyroiditis	–
Ajilisa OA et al. (2023)^27	DDTI	980	658 benign, 322 malignant	–
Zhiqiang Zheng et al. (2023)^28	Inner Mongolia People's Hospital (October 2017–December 2020)	4021	Benign nodules: 1,844 imagesMalignant nodules: 2,177 images	GE LOGIQ E9
S. Menderes and I. Sahin (2024)^29	DDTI	440	Benign: 59 images, Malignant: 261 images, Normal: 100 images	–
Na Feng et al. (2024)^30	Five hospitals in China between October 2019 and October 2022	25,298	Nodules: 9,649 thyroid nodules. Patients: 8,455 individuals. Only nodules less than 1 cm were included.	GE, Siemens, Toshiba, Mindray
Veda Reddy et al. (2024)^31	Kaggle & UCI	594	Thyroid Cancer - 99, Hyperthyroid – 77, Hypothyroid – 110Thyroid Nodules – 146Normal Thyroid – 88Thyroiditis – 74	–
Jatinder Kumar et al. (2024)^32	DDTI	427	from 299 individuals, 357 malignant, 70 benign	–
Sheeja Agustin et al. (2024)^33	Thyroid Digital Image Database (TDID)	450	298 patients: benign – 52, malignant – 246	–
P. Sanju et al. (2025)^34	UCI	10,000	5,000 benign, 5,000 malignant	–
Yan Xu et al. (2025)^35	Zhejiang Rongjun Hospital, China.	1134	630 patients, 589 benign, 545 malignant	Mylab90 (Esaote, Italy)Philips EPIQ5 and EPIQ7 (Netherlands)5–12 MHz transducers
Rafee SM et al. (2025)^36	Kaggle	2,450	Benign, malignant, or normal thyroid nodules	–
Zeeshan Ali Haider et al. (2025)^37	Kaggle & local hospitals	13389	12 thyroid disease categories	–
Yasaman Sharif et al. (2025)^38	PERSIAN cohort study	2450	1037 patients between 2018 and 2020	Philips Affinity 50G Ultrasound Machine with a 12.5 MHz frequency and a 5 cm linear transducer
Yang C et al. (2025)^39	Multiple hospitals (specific names not disclosed)	600	Grayscale images	GE Logiq E9 with 7–12 MHz frequency and GE 12L-RS linear array transducer
Hassan S. Al Khamsan et al. (2025)^40	Ultrasoundcases.info	1137	594 benign543 malignant	–
Hui Bi et al. (2025)^41	1: National Cancer Institute (NCI) and the American Association of Physicists in Medicine (AAPM)2. The Second People's Hospital of Changzhou.	1: 4382: 1096	Diverse dataset	Philips Healthcare (Netherlands)Siemens Healthineers (Germany)GE Healthcare (USA)

Pre-processing techniques

The table summarizes the wide array of image preprocessing techniques employed in different studies between 2013 and 2025. Common methods included image resizing, conversion to grayscale, normalization, cropping, or selection of the region of interest (ROI), and data augmentation to enhance dataset quality and model performance. Most studies also employed noise removal and filtering techniques such as the median, Gaussian, bilateral, and anisotropic diffusion filters to make the image more readable. Histogram equalization and contrast stretching were also used to further improve image quality, while morphological operations were used to create structural refinement. Some advanced studies employed methods such as neutrosophic transformation, active contour without edge (ACWE), feature-wise centering, HSV transformation, and expert-based manual annotation to achieve higher preprocessing accuracy. Overall, the evolution exhibits a continuing trend toward powerful and customized preprocessing chains for improved image-based model performance.

Feature extraction techniques

Table 2 summarizes top thyroid ultrasound image analysis papers from 2013–2025 on feature extraction methods employed. The listed papers demonstrate the transition from earlier hand-crafted descriptors such as statistical, textual, and transform-domain ones to later automatically learning discriminatory representations through convolutional and transformer networks. The listing includes the dominant feature extraction method(s), feature selection/enhancement (if any), employed classifier or network, and top results published. This comparative snapshot demonstrates how feature extraction evolved to support more automatic and accurate thyroid nodule and Hashimoto's thyroiditis diagnosis.

Table 2.

Feature extraction techniques.

Year and authors	Feature extraction technique(s)	Feature selection/enhancement	Classifier/network used	Notable outcomes
Acharya et al. (2013)^7	Shannon Entropy, Gabor Wavelets, Hu's Moments, Fractal Dimension, Hurst Exponent, Higher-Order Spectra	–	Not specified (focus on feature extraction)	Captured both texture and frequency info for Hashimoto's classification
Acharya et al. (2014)^8	Stationary Wavelet Transform (SWT)	–	–	Multi-resolution texture features for thyroid classification
Omiotek et al. (2015)^9	Histogram, GLCM, RLM, Haar Wavelet	–	Decision Trees, MLP	Texture info effective for Hashimoto's classification
Ardakani et al. (2015)^10	Histogram, Gradient, RLM, GLCM, Wavelet	PCA, LDA, NDA	ANN	Combined handcrafted + dimensionality reduction improved classification
Koundal et al. (2016)^11	Spatial Neutrosophic L-Means (SNLM) segmentation	SNDRLS enhancement	–	Clear boundary definition but no classification
Xie et al. (2016)^43	Manual feature extraction based on BTA U1–U5 sonographic criteria	–	–	Integrated clinical sonographic features
Nugroho et al. (2017)^12	GLCM, GLRLM, Histograms	–	MLP	Distinction between solid vs. cystic nodules
Xia et al. (2017)^13	Manual features + ReliefF selection	ReliefF	Extreme Learning Machine (ELM)	Fast prediction after selection
Sun et al. (2018)^14	FCN-AlexNet (transfer learning)	–	CNN (deep learning)	Automatic extraction of high-level features
Frannita et al. (2018)^15	Lacunarity + GLCM texture	–	MLP	Captured irregularities; high performance
Cen et al. (2019)^16	Pretrained CNN (transfer learning)	Retrained FC layers	Softmax	Robust nodule classification
Liu et al. (2019)^17	Multi-scale, multi-branch CNN guided by clinical features	–	CNN	Better detection of small nodules
Liu et al. (2020)^19	ResNet-34 as deep feature extractor in U-Net	–	Deep learning segmentation	Enhanced segmentation accuracy
Chen et al. (2020)^44	GLCM, fuzzy histograms, DCT vs CNN	–	ANN, CNN	Compared traditional vs CNN accuracy
Shin et al. (2020)^20	GLCM, GLRLM	–	ANN, SVM	Limited automation despite high specificity
Vadhiraj et al. (2021)^21	GLCM	–	SVM, ANN	SVM > ANN for sensitivity
Sharifi et al. (2021)^45	Pretrained CNN (VGGNet, ResNet, Inception)	Optional augmentation	Random Forest/SVM	Malignancy prediction with pre-trained features
Wang et al. (2021)^22	ResNet34 segmentation + FC ResNet50 classification	Semi-supervised learning	CNN	Leveraged labeled and unlabeled data
Vasile et al. (2021)^23	5 CNNs + VGG-19 ensemble	Probabilistic fusion	–	Enhanced diagnostic accuracy
Pavithra et al. (2022)^24	CNNs (AlexNet, ResNet-50, VGG-19) with transfer learning	–	CNN	Pretrained convolutional layers used for features
Srivastava and Kumar (2022)^25	CNN + Grid Search Optimization (GSO-CNN)	Hyperparameter tuning	CNN	Optimized performance
Yao et al. (2023)^26	HT-Net, ThyNet (CNN)	Big data seronegative AIT	CNN	Automatic feature extraction for large datasets
Ajilisa et al. (2023)^27	Inception + SE modules (multi-level)	Multi-level transfer learning	–	Better generalization on small datasets
Zheng et al. (2023)^28	Dual-branch CNN (ResNet50/101 + GSAM)	–	CNN	Enhanced sensitivity and interpretability
Sahin and Menderes (2024)^29	CNN-based deep learning	–	SVM and Random Forest for comparison	Summarized CNN-based methods
Na Feng et al. (2024)^30	EfficientNet-B0 pre-trained conv layers (<1 cm nodules)	–	Dense layer (Sigmoid)	Effective small-nodule classification
Veda Reddy et al. (2024)^31	Vision Transformer (ViT), EfficientNetB3, Inception ResNet V2	–	–	ViT attention for important regions
Rajeswari V et al. (2024)^42	Gabor Wavelet Transform + En-SwinT +	Hybrid (traditional + DL)	–	Multi-scale abstraction and high precision
Jatinder Kumar et al. (2024)^32	ResNet-18, VGG-19, AlexNet (transfer learning)	Mid-CNN layers	–	Classification on DDTI dataset
Sheeja Agustin et al. (2024)^33	GLCM + MbPSRA	–	Residual U-Net	High sensitivity and specificity
Sanju et al. (2025)^34	CNN + Hybrid Feature Selection	Random Forest	–	Improved efficiency and relevance
Xu Y et al. (2025)^35	PyRadiomics handcrafted + deep CNN (VGG16, ResNet50, InceptionV3)	Dual-stream learning	–	Higher interpretability via manual ROIs
Rafee S et al. (2025)^36	Texture, edge, shape, histogram features	CNN + SVM + PSO optimization	–	Improved diagnostic potential
Ali Zeeshan et al. (2025)^37	InceptionV3 + personalized ResNet	Dense Softmax (Adam/SGD)	–	Enhanced classification
Sharifi et al. (2025)^38	Faster R-CNN + ResNet-101 for detection; Xception for discrimination	–	ACR-TIRADS risk stratification	21 nodular features
Yang C et al. (2025)^39	Swin U-Net (self-attention + multi-scale convolution)	–	–	Robust small/blurry nodule detection
Hassan S et al. (2025)^40	EfficientNetB0 + parallel personalized CNN	ADF preprocessing	–	Better thyroid tumor classification
Hui Bi et al. (2025)^41	UMAM-NET (ResNet50 + Multi-Subnet Attention)	Spatially attentive features	–	Classify levels of malignancy

Table 2 works exhibit a general development trend from hand-crafted features to hybrid and deep learning–based feature representations. Earlier work greatly relied on texture and frequency-domain descriptors and traditional classifiers (ANN, SVM, MLP), while recent works employ pre-trained CNNs, multi-branch networks, and transformers (ViT, Swin) to learn multi-scale and attention-guided features. Hybrid approaches – with both hand-designed descriptors, radiomics, and deep features – are gaining popularity as means to promote interpretability and performance, particularly for small or indeterminate nodules. This trend indicates that future work will continue to integrate domain knowledge with advanced architectures for more explainable and precise thyroid nodule assessment.

The evolution of feature extraction techniques in thyroid ultrasound analysis (2013–2025)

• -

  Handcrafted texture features (2013–2017) – Entropy/Gabor/Wavelets/Histograms/GLCM/RLM, Classical ML: ANN, SVM, Decision Trees, MLP.

• -

  Hybrid features + Dimensionality reduction (2015–2018) – PCA, LDA, NDA + ReliefF, ELM, ANN.

• -

  Early deep learning and transfer learning (2018–2020) -–AlexNet, FCN, ResNet, CNN as automatic feature extractor.

• -

  Advanced CNN architectures (2021–2023) – Multi-branch, dual-branch, ensemble CNNs, Clinical knowledge + semi-supervised learning.

• -

  Hybrid and transformer-based models (2023–2025) – ViT, Swin U-Net, En-SwinT + , EfficientNet, CNN + SVM/Random Forest, Hybrid handcrafted + deep features, Attention mechanisms and multi-scale features.

Classification techniques

Table 3 offers comparative overview of major classification techniques employed in thyroid nodule/Hashimoto's thyroiditis image processing from 2013 to 2025. It simplifies classifiers taken into consideration by one, marks out essential strengths achieved, and counts restrictions stated. Owing to it, readers will take pleasure in browsing at once how steadily old hand-designed feature classifiers (e.g., SVM, k-NN, Decision Trees) were overhauled by deep learning and hybrid ones (e.g., CNNs, Transformers, ensembles), how these reforms have been followed by increasing data needs and computational complexity.

Table 3.

Classification techniques.

Year and authors	Classifier(s) used	Strengths	Weaknesses
Acharya et al. (2013)^7	SVM (linear, polynomial, RBF), KNN, RBF Neural Network	Multiple classifiers tested; polynomial SVM highest accuracy	Scanner-dependent results
Acharya et al. (2014)^8	SVM, Decision Tree, Fuzzy Classifier, KNN	SWT features support multi-resolution; fuzzy achieved best	Sensitivity to feature extraction quality
Omiotek et al. (2015)^9	Decision Tree (C&RT), MLP	Decision Tree higher specificity; MLP high accuracy	MLP prone to false positives
Ardakani et al. (2015)^10	ANN (after PCA/LDA/NDA)	Dimension reduction boosted ANN performance; high specificity	Requires careful preprocessing
Nugroho et al. (2017)^12	MLP	Simple architecture; works on texture	Very small dataset; manual ROI
Xia et al. (2017)^13	Extreme Learning Machine (ELM) with ReliefF	Fast learning; attribute selection improved accuracy	Subjective feature extraction
Sun et al. (2018)^14	FCN + AlexNet (Transfer Learning)	Deep features improved malignancy detection; end-to-end	Computational cost
Frannita et al. (2018)^15	MLP	High diagnostic precision with lacunarity features	Small dataset, needs meticulous preprocessing
Cen et al. (2019)^16	Softmax on CNN features	Larger datasets improved performance	Transfer learning dependent
Liu et al. (2019)^17	Multi-branch CNN guided by clinical knowledge	Domain knowledge integration boosted accuracy	Complexity of model
Chen et al. (2020)^44	SVM vs CNN (Hybrid)	CNN robust and accurate	Hybrid model complexity
Shin et al. (2020)^20	ANN and SVM	Outperformed human specificity	Low sensitivity, manual segmentation
Vadhiraj et al. (2021)^21	SVM, ANN	SVM high accuracy with GLCM features	ANN poor sensitivity
Sharifi et al. (2021)^45	CNNs, optional SVM and RF	Mostly automated feature learning	Needs large annotated data
Vasile et al. (2021)^23	Ensemble of 5 CNNs + VGG-19	Probability averaging improved generalizability	High computational cost
Pavithra et al. (2022)^24	AlexNet, ResNet-50, VGG-19, customized CNN	Customized CNN very high accuracy	Heavy computing resource requirement
Srivastava and Kumar (2022)^25	GSO-CNN with tuning	Grid search improved small DB performance	Segmentation/enhancement critical
Yao et al. (2023)^26	HT-Net, ThyNet (CNNs)	Automatic feature learning	Huge annotated databases needed
Ajilisa et al. (2023)^27	Updated Inception + SE modules	Strong generalization with multi-level transfer learning	High compute
Zheng et al. (2023)^28	Dual-branch ResNet50/101 + GSAM	Combined global and local features	High computing resource cost
Sahin and Menderes (2024)^29	SVM, Random Forest, k-NN, CNNs	CNNs outperformed; hybrid methods stressed	Real-world testing needed
Na Feng et al. (2024)^30	EfficientNet-B0 + Binary Sigmoid Classifier	Good for small nodules	Limited to binary classification
Veda Reddy et al. (2024)^31	ViT, EfficientNetB3, Inception ResNetV2	Attention to relevant image parts	Training complexity
Rajeswari V et al. (2024)^42	Enhanced Swin Transformer (En-SwinT + )	Ontology-based classification improved accuracy	Slow due to deep model fusion
Jatinder Kumar et al. (2024)^32	ResNet-18, VGG-19, AlexNet (TL)	Transfer learning effective	Performance variability
Sheeja Agustin et al. (2024)^33	Residual U-Net + MbPSRA	Low false positives	Model complexity
Sanju et al. (2025)^34	CNN + Random Forest (HFS-DLF)	Combined deep and traditional learning	Needs careful framework tuning
Xu Y. et al. (2025)^35	InceptionV3 + SVM hybrid	Combining deep + handcrafted features	Lower AUC vs others
Rafee S. et al. (2025)^36	CNN, CNN + SVM, PSO-CNN	Minimum execution time	PSO optimization complexity
Ali Zeeshan et al. (2025)^37	InceptionV3 + adapted ResNet + Dense Softmax	Mixed optimizers improved training	High training time
Sharifi et al. (2025)^38	Faster R-CNN (ResNet-101) + Xception	Higher precision than radiologists	Complex detection-classification setup
Hassan S. et al. (2025)^40	EfficientNetB0 + anisotropic diffusion filter + custom CNN	Image enhancement improved accuracy	Hybrid model tuning
Hui Bi et al. (2025)^41	UMAM-NET (ResNet50 + MN-ATT)	Interpretability enhanced via spatial attention	More training data required

For these examined studies, one distinct trend emerges: early work (2013–2015) employed predominantly SVM, Decision Trees and MLPs upon manually-designed texture or wavelet features, achieving modest accuracies yet suffering from variability through scanners and manual region selection; from 2017 to 2020 scientists began to employ CNNs, extreme learning machines and hybrid SVM-CNN processing extensively, achieving increases in accuracies frequently in excess of 90%, though frequently at larger computational cost; and from 2021 to 2025 ensembling, Transformers (ViT, SwinT), specialised CNNs and hybrid systems using clinical knowledge or image enhancement have gained popularity, achieving very high accuracies and specificity (often over 95–99%) with improved generalizability though at massive annotated dataset requirement, stringent tenability and huge computing resources. Taken as whole, this trend signifies shift from manually-designed features through designers to fully-automatic deep and hybrid models, both exhibiting resultant improvement in terms of performances as well as practical limitation remaining prior to realistic deployment.

The advancement of thyroid nodule classification shows an evident shift away from traditional machine learning methods like SVM, KNN, Decision Trees, and ANN (2013–2017) with moderate accuracies (80–90%), and in favor of deep learning methods (2018+) using CNNs, transfer learning, and ensemble techniques, substantially enhancing performance (oftentimes>95%). Early methods heavily depended on feature engineering and were limited with small datasets, while deep learning methods as AlexNet, ResNet, VGG, Inception, EfficientNet, ViT, Swin Transformers, and hybrid CNN-based systems attained automated feature learning, robustness, and scale, with some methods achieving near-perfect (>99%) accuracies. The most recent trends are the hybrid methods combining deep and hand-engineered features, ensemble architecture, transformers, attention mechanisms, and clinical integration, with higher accuracies, sensitivity, and interpretability, although there are challenges concerning computational cost, dataset size, and field testing.

Results and discussions

The literature surveyed from 2013 through 2025 exhibits a clear advancement in precision and efficacy of automatic thyroid nodule classification through ultrasound imaging. Initial investigations heavily depended upon conventional machine learning classifiers such as SVM, KNN, and Decision Trees with manually engineered features and obtained a fair level of performance (approximately 80–90% accuracy). Feature extraction techniques gradually became more varied and advanced over time, such as wavelet transforms, GLCM, and higher-order spectra. Hybrid models and ANN-based systems arrived around 2015 and drove accuracies above 97%, though several still suffered due to factors such as limited datasets and excessive false-positives. Segmentation tasks around 2016–2018 also experienced a very good level of Dice coefficient (>90%) through innovative clustering and FCN-based designs and represent a valuable step towards proper localization.

From 2019 onward, deep learning models, particularly CNNs, transfer learning architectures, and attention mechanisms (e.g., ViT, Swin Transformer), became the dominant approach. These methods consistently achieved high classification accuracies – often exceeding 95% – with some models reaching as high as 99% accuracy and AUC values of 0.99. More recent works incorporated ensemble models, semi-supervised learning, and hybrid approaches to overcome challenges like small dataset sizes, edge detection issues, and interpretability. While models like EfficientNet, ResNet, and custom CNNs yielded impressive results, interpretability, generalizability across institutions, and computational complexity remain persistent challenges. Overall, the field has shown significant advancement, suggesting strong clinical potential pending further real-world validation and standardization. The final results achieved in different studies are analyzed in detail in the Table 4.

Table 4.

Result analysis.

Year	Study	Method/model	Accuracy	Sensitivity	Specificity	AUC/F1/dice/other
2013	Acharya et al.^7	SVM (Polynomial)	80%	76%	84%	83.3% PPV
2014	Acharya et al.^8	Fuzzy Classifier	84.6%	–	–	–
2015	Z. Omiotek et al.^9	Decision Tree and MLP	89.41%	–	–	–
2015	A. A. Ardakani et al.^10	ANN	97.14%	94.45%	100%	–
2016	DeepikaKoundal et al.^11	SNDRLS	–	95.4% TPR	–	94.2% Dice
2016	Cheng Xie et al.^43	BTA Review	–	–	–	No quant. Results
2017	Nugroho et al.^12	MLP	89.74%	–	–	–
2017	Xia et al.^13	ELM	87.72%	–	–	–
2018	Sun et al.^14	FCN-AlexNet	90.79%	–	–	0.9516 AUC
2018	Frannita et al.^15	MLP	98.97%	–	–	–
2019	Xueya Cen et al.^16	CNN + Transfer Learning	89.5%	–	–	–
2019	Tianjiao Liu et al.^17	Clinical CNN	97.1%/94.9%	–	–	–
2020	Mengya Liu et al.^19	Optimized U-Net	–	–	–	89.5% Dice
2020	Hervé Monpeyssen et al.^18	HIFU	–	–	–	30–35% volume reduction
2020	Junying Chen et al.^44	Hybrid DL	High	–	–	Best performance noted
2020	Ilah Shin et al.^20	ANN	Improved	–	Improved	–
2021	Vadhiraj et al.^21	SVM/ANN	96%/74.57%	–/49.14%	100%/–	–
2021	YasamanSharifi et al.^45	CNN Models	99.13%	99.7%	–	0.99 AUC
2021	Qingsong Wang et al.^22	ResNet34 + FCN ResNet50	–	ResNet34 + FCN ResNet50	–	89.7% Dice, 93.2% Precision
2021	CorinaVasile et al.^23	Ensemble CNN + VGG19	97.35%	95.75%	98.43%	0.96 AUC
2022	S. Pavithra et al.^24	Custom CNN	99.17%	–	–	–
2022	RajshreeSrivastava et al.^25	GSO-CNN	96.02%	–	–	–
2023	Shengsheng Yao et al.^26	HT-Net	Improved	–	–	No values given
2023	Ajilisa O.A et al.^27	Inception + SE + TL	92.77%	–	–	0.9537 AUC
2023	ZhiqiangZheng et al.^28	Dual-branch CNN	86.34%	90.48%	81.29%	–
2024	Sahin S and Menderes I.^29	DL Approaches	>90%	–	–	–
2024	Na Feng et al.^30	EfficientNet-B0	87.1%	86.5%	87.6%	0.931 AUC
2024	Veda Reddy et al.^31	ViT + EfficientNetB3	95.4%	–	–	0.94 AUC
2024	Rajeswari V et al.^42	En-SwinT+	99.45%	–	–	–
2024	Jatinder Kumar et al.^32	ResNet-18	97.13%	–	–	–
2024	Sheeja Agustin et al.^33	MbPSRA + Residual U-Net	95%	97%	98%	94% detection rate
2025	P. Sanju et al.^34	HFS-DLF (CNN + RF)	96.3%	96.5%	–	96.15% F1, 95.8% Precision
2025	Xu Y. et al.^39	InceptionV3 + SVM	–	–	–	0.783 AUC
2025	Rafee S. et al.^36	PSO-CNN	94%	92%	98%	96% F1, 160ms/img
2025	Ali Zeeshan et al.^37	Modified ResNet + Inception	97%	–	–	High PR/F1
2025	YasamanSharifi et al.^38	Faster R-CNN + Xception	98%	–	–	0.99 AUC
2025	Yang C et al.^39	Swin U-Net	–	–	–	0.78 Dice
2025	Hassan S et al.^40	EfficientNetB0 + CNN	92.28%	–	–	92.76% F1
2025	Hui Bi et al.^41	UMAM-NET	High	–	–	High F1, no values given

Limitations

The limitations mentioned by Acharya et al.^7,8 were a limited dataset size and machine dependency, and a narrow generalizability that does not extend beyond population-specific examples. Omiotek et al.⁹ and Ardakani et al.¹⁰ were limited by dataset size and false-positive rates or a narrow set of feature engineering tools. Koundal et al.¹¹ were limited by a lack of classification and external validation, while Xie et al.⁴³ were limited by subjective interpretation and manual rating.

Nugroho et al.¹² and Xia et al.¹³ were confronted with limited available data and a need to depend on manual feature extraction or selecting regions of interest (ROI). Sun et al.¹⁴ were faced with the risk of misidentification and computational load, and Frannita et al.¹⁵ were confronted with dataset size and complex pre-processing. Cen et al.¹⁶ emphasized sample-rich datasets, and Liu et al.¹⁷ were confronted with model complexity and class imbalance. 2020 articles by Chen⁴⁴ and Shin²⁰ pointed out computational demand, dependence on massive datasets, and shortage of external validation. Monpeyssen et al.¹⁸ detected modest effectiveness and technological constraints in the therapies of HIFU. Vadhiraj et al.²¹ indicated low ANN sensitivity, and Sharifi et al.⁴⁵ pointed out model interpretability and a lack.

Wang et al.²² cited training complexity and label quality issues, whereas Vasile et al.²³ cited no such limitation. Pavithra et al.²⁴ and Srivastava et al.²⁵ faced over fitting and high computational load along with stringent pre-processing. Yao et al.²⁶ and Ajilisa et al.²⁷ cited a need for annotated data as well as domain adaptation. Zheng et al.²⁸ corrected for limited generalizability using single-site data. Sahin and Menderes²⁹ pointed out a lack of standard datasets and real-world validation. Feng et al.³⁰ called for increased clinical testing. Reddy et al.³¹ faced widespread training expense and data diversity necessities, and Rajeswari et al.⁴² pointed out the volume of data and computing intensity.

Kumar et al.³² acknowledged data imbalance and interpretability issues. Agustin et al.³³ found no explicit gap mentioned. Sanju et al.³⁴ faced model interpretability and computationally intensive issues. Xu et al.³⁵ required multicenter verification due to a small sample of data. Rafee et al.³⁶ and Ali et al.³⁷ had not highlighted and identified the study limitations. Sharifi et al.³⁸ emphasized multi-institutional data and augmentation. Yang et al.³⁹ addressed the performance of segmentation in the case of small or blurry nodules. Hassan et al.⁴⁰ indicated issues of a small dataset and model generalizability, and Bi et al.⁴¹ indicated complexity and data limitation. The overall comparison of various techniques and its results are given in Table 5.

Table 5.

Review of overall techniques and results achieved.

Study	Objective	Dataset	Methods used	Features extracted	Classifier/segmentation	Accuracy/metrics
Selvathi and Sharnitha, (2011)^6	Classification and segmentation of thyroid gland in US images	Dataset from online sources and radiologists (image count not specified)	SVM and ELM classifiers for classification and segmentation	Mean, Variance, Coefficient of Local Variation, Histogram, NMSID, Homogeneity	SVM for classification, ELM for segmentation	ELM segmentation more accurate; SVM classification accuracy: 96.7%
Acharya et al., (2013)^7	Diagnosis of Hashimoto's thyroiditis using CAD system	200 images (100 normal, 100 HT-affected)	CAD system with entropy, Gabor, moments, HOS analysis	12 features (Entropy, Gabor wavelet, Hu's moments, Fractal dimension, HOS)	SVM, KNN, RBPNN	Best accuracy: 80% (SVM), Sens: 76%, Spec: 84%, PPV: 83.3%
Acharya et al., (2014)^8	CAD system (ThyroScan) for Hashimoto's thyroiditis	526 images (232 normal, 294 HT-affected)	Stationary Wavelet Transform + T-test for feature selection	7 features based on SWT (Stationary Wavelet Transform)	SVM, KNN, Fuzzy Classifier, Decision Tree	Best: 84.6% accuracy (Fuzzy Classifier), Sens: 82.8%, Spec: 87.0%, PPV: 88.9%
Abbasian et al. (2015)^10	Evaluate texture analysis in CAD systems	70 ROIs (26 benign, 44 malignant) Accuvix V20 ultrasound, 5–13 MHz	POE + ACC, Fisher, PCA, LDA, NDA	GLCM, Run-Length, Histogram, AR model, Wavelets	NDA + ANN, 1-NN + PCA/LDA	Az = 0.9722, Sens: 94.45%, Spec: 100%, Acc: 97.14%
Koundal et al. (2016)^11	Delineate thyroid nodules using unsupervised methods	42 US images from PGIMER, India IU22 Philips X Matrix, 17.5 MHz	Spatial Neutrosophic Clustering + Level Set (SNDRLS)	Not texture features but segmentation metrics	SNLM + SNDRLS (automated)	Dice: 94.25%, Overlap: 93.15%, TPR: 95.45%, FPR: 7.32%
Nugroho et al. (2017)^12	Classify cystic vs. solid nodules	39 images (25 cystic, 14 solid) from Sardjito Hospital	Median filter, Morphological ops, Histogram Equalization, ACWE	Histogram, GLCM, GLRLM (33 features)	Multilayer Perceptron (MLP)	Acc: 89.74%, Sens: 88.89%, Spec: 91.67%, PPV: 96%, NPV: 78.57%
Xia et al. (2017)^13	Differentiate benign and malignant nodules	203 nodules (114 benign, 89 malignant) from Wenzhou Central Hospital	ReliefF-based feature selection; Extreme Learning Machine (ELM)	Shape, Echogenicity, Margin, Calcification, Composition	Classifier: ELM	Accuracy: 87.72% Sensitivity: 78.89% Specificity: 94.55% AUC: 0.8672
Sun et al. (2018)^14	Lesion localization and benign/malignant classification	Not specified (external test sets used)	FCN-AlexNet for segmentation; Transfer Learning with AlexNet for classification	Deep learned features (from AlexNet)	FCN for segmentation; AlexNet for classification	Outperformed TBD, RGI, PAORGB, ASPS methods in IoU and AUC;
Frannita et al. (2018)^15	Classify nodules into solid, cystic, complex	97 US images (34 solid, 32 cystic, 31 complex); Yogyakarta, Indonesia	Preprocessing, Morphological ops, Active Contour Segmentation	Histogram, GLCM, GLRLM, Lacunarity	Classifier: MLP	Accuracy: 98.97% Sensitivity: 98.92% Specificity: 99.47% PPV: 99.05% NPV: 99.50%
Liu et al. (2019)^17	Automated detection and classification using domain knowledge	Dataset I: 7690 images (5139 malignant, 2551 benign) Dataset II: 450 images (322 malignant, 128 benign)	Multi-scale detection (FPN + Faster R-CNN); Multi-branch classification network	Multi-view semantic features + clinical characteristics (e.g., margin, echo, calcification)	CNN (Faster R-CNN for detection; Multi-branch CNN for classification)	Detection Accuracy: 97.5% Classification Accuracy: 97.1% Better than radiologists on Dataset II by 8% margin
Liu et al., (2020)^19	End-to-end thyroid nodule segmentation	544 ultrasound images	Optimized U-Net (with ResNet-34 backbone), Test-Time Augmentation (TTA)	Image-level features learned by CNN	U-Net + ResNet-34 with dense skip connections	Dice Coefficient: 89.5%
Shin et al., (2020)^20	Differentiate follicular adenoma from carcinoma via ML	348 nodules (252 adenoma, 96 carcinoma)	Manual segmentation, LASSO for feature selection, ANN and SVM classifiers	96 radiomic features (GLCM, GLRLM, Gabor, Haar, shape) → 10 selected	ANN, SVM	ANN: Acc. 74.1%, Sens. 32.3%, Spec. 90.1% SVM: Acc. 69.0%, Sens. 41.7%, Spec. 79.4%
Monpeyssen et al., (2020)^18	Assess long-term efficacy of HIFU for benign nodules	65 patients (single nodule each), 3-year follow-up	High-intensity Focused Ultrasound (HIFU)	Nodule volume (reduction over time)	None (Therapy evaluation, not segmentation)	Volume Reduction: ∼31.9% at 36 months, Tech. Efficacy >50% in ∼7.4%
Chen et al., 2020 ^44	Survey and analyze segmentation methods for thyroid ultrasound	Multiple datasets, including 22123 images for CNN	Contour, region-based, machine learning (SVM, DT, KNN), deep learning (CNN, U-Net)	Texture (GLCM, GLRLM), histogram, shape, wavelet	Multiple (CNN, CCNN, SVM, etc.)	CNN: Dice 92.24%, CCNN: MO 87.00%, SVM: Acc. 91.30%
Sargolzaei et al. (2021)^45	Systematic review of DL methods for thyroid nodule analysis in US images	Multiple datasets (public/local); Image counts vary from 300–10000	Various DL architectures including VGG, GoogLeNet, Inception, CNN, SVM	Traditional (HOG, LBP) and DL features, some used TI-RADS, ROI features	VGG, Inception-v3, GoogLeNet, SVM, Random Forest, CNN	Accuracy up to 99.13%, Sensitivity up to 100%, AUC up to 99.7%
Pavithra et al. (2022)^24	Classification of thyroid nodules into benign and malignant using CNN and transfer learning	Public dataset (TI-RADS labeled); exact image count not specified	CNN trained from scratch; compared with AlexNet, VGG-19, ResNet-50	End-to-end learned deep features from US images	CNN, AlexNet, VGG-19, ResNet-50	CNN Accuracy: 99.17%, Sensitivity: 0.98, Precision: 0.97, F1 Score: 1.0
Zheng et al. (2023)^28	Propose a novel diagnostic model using multiscale localization and dual-branch classification network	4021 US images from Inner Mongolia People's Hospital (GE LOGIQ E9)	Multiscale Faster RCNN + FPN for localization; GSAM for attention-based classification	Low and high-level deep features, aspect ratio of nodules, fused via attention	Multiscale localization + dual-branch classification with ResNet + GSAM	Classification Accuracy: 86.34%, Sensitivity: 90.48%, Specificity: 81.29%
Wang et al., (2021)^22	Automatic detection of thyroid nodules using semi-supervised learning	2032 training images, 400 test images (US)	Improved U-Net with ResNet34 backbone for segmentation, ResNet50 for classification, semi-supervised learning with pseudo-labels	CNN-based features via ResNet	U-Net (segmentation), FCN-ResNet50 (classification)	Dice: 89.7%, Precision: 93.2%, Recall: 96%
Vasile et al., (2021)^23	Intelligent diagnosis of thyroid US using deep learning ensemble	2797 US images (autoimmune, micro-nodular, nodular, normal)	Ensemble of end-to-end trained 5-CNN and fine-tuned VGG-19	Learned deep features from CNN and VGG	5-CNN, VGG-19, Ensemble CNN-VGG	Ensemble Accuracy: 97.35%, Sens: 95.75%, Spec: 98.43%, PPV: 95.41%, NPV: 98.05%, AUC: 0.96
Vadhiraj et al., (2021)^21	Benign–malignant classification of thyroid nodules using ML	99 patients (33 benign, 66 malignant)	Median filter, image binarization, GLCM, GUI, MIL framework	7 GLCM features (Energy, Entropy, Correlation, etc.)	ANN and SVM	SVM: 96% Accuracy, ANN: 75%; SVM showed higher sensitivity and specificity
Rajshree Srivastava and Pardeep Kumar (2022)^25	Identification and classification of thyroid nodules using optimized deep learning	TDID: 295 images (public) Local: 654 images (from Kriti Scanning Center, India)	GSO-CNN model Preprocessing, Segmentation, Morphological ops	Morphological features Gaussian blur-enhanced regions	CNN (with Grid Search Optimization), segmentation via boundary detection	Accuracy: 95.30% (TDID), 96.02% (Local) Sensitivity: 96.66%, 96.70% Specificity: 94.87%, 95% F1-Score: 97.20%, 98.34%
Ajilisa et al. (2023)^27	Differentiate benign and malignant thyroid nodules using improved CNN and multi-level transfer learning	Private dataset (not explicitly named), transfer learning with breast ultrasound dataset	Improved Inception Network Multi-level Transfer Learning Squeeze and Excitation modules	Automatically learned deep features via inception-SE modules	CNN (Improved Inception + SE), no separate segmentation phase	Precision: 0.9057 (benign), 0.9667 (malignant) Recall: 0.9796 (benign), 0.8529 (malignant) F1-score: 0.9412 (benign), 0.9062 (malignant) AUC: 0.9537
Kumar et al. (2024)^32	Assess deep learning models (ResNet-18, VGG-19, AlexNet) for classifying thyroid nodules from ultrasound images.	Digital Database of Thyroid Ultrasound Images (DDTI)	Pretrained DL models: ResNet-18, VGG-19, AlexNet	DL-based features from ultrasound images	ResNet-18, VGG-19, AlexNet (fine-tuned)	ResNet-18: 97.13%, VGG-19: 90.31%, AlexNet: 83.59%
Reddy et al. (2024)^31	Use Vision Transformers (ViT) for thyroid abnormality detection and classification.	Custom dataset (Ultrasound, CT, MRI images)	Vision Transformer (ViT), EfficientNetB3, Inception ResNet V2	Patch embeddings, ViT outputs, deep features	ViT with attention mechanisms, model fusion	High accuracy reported; exact metric values not specified
Sanju et al. (2025)^34	Propose Hybrid Feature Selection and DL Framework (HFSDLF) for thyroid disease prediction.	UCI Thyroid Dataset	Random Forest, PCA, L1 Regularization, Hybrid framework	Selected features via PCA + L1, RF-based selection	Optimized Random Forest	Accuracy: 96.3%; superior to DT and LR
Xu et al. (2025)^35	Develop fusion model using traditional ML and deep transfer learning for thyroid nodule classification.	1134 ultrasound images (589 benign, 545 malignant)	LASSO, PCA, SVM, Inception V3, model fusion	Radiomic + deep features (PCA-reduced)	Fusion of Inception V3 + SVM	Fusion model AUC: 0.783; SVM AUC: 0.748; Inception V3 AUC: 0.763
Rafee et al. (2025)^36	Optimize CNNs using Particle Swarm Optimization (PSO) for thyroid medical image prediction.	Large thyroid image dataset (unspecified size)	CNN, PSO for hyperparameter optimization	Edge, texture, CNN-learned features	PSO-optimized CNN architecture	Improved accuracy, sensitivity, specificity; exact values not stated
Menderes and Şahin (2024)^29	Evaluate ML and DL models for classifying thyroid nodules as benign, malignant, or normal.	DDTI Dataset (375 cases, 440 images)	Inception-ResNet-v2, ResNet, data augmentation	CNN-extracted features (image-based); TIRADS categories	CNN (ResNet variants), logistic regression, LSTM, DMGSO, SVM, Random Forest	Best model: High classification accuracy; AUC and PPV emphasized; improvement in benign class noted
Na Feng et al. (2024)^30	Develop TNT-Net for diagnosing nodules < 1 cm using dual-channel ultrasound images.	9649 nodules from 8455 patients (5 hospitals)	Dual-channel Swin Transformer (TNT-Net)	Transverse + longitudinal US features via Swin Transformer	TNT-Net (dual-channel Swin Transformer)	AUC: 0.953 (internal), 0.941 (external); Outperforms DCNNs (ResNet50, DenseNet121, Inception V3)
Hariharan et al. (2024)^42	Integrate ontology matching with DL for improved thyroid disease detection.	Mixed dataset: 594 images (UCI + hospital; various diseases)	IWSO optimization, Gabor Wavelet Transform, En-SwinT + model	GWT-extracted wavelet features; enhanced via ontology matching	En-SwinT + (enhanced Swin Transformer)	Accuracy: 99.3%; claimed superior to other methods
Agustin et al. (2024)^33	Segment and classify thyroid nodules from ultrasound images.	TDID Dataset (298 patients: 41 + 11 benign, 196 + 50 malignant)	ADF preprocessing, bilateral filter segmentation, GLCM + MbPSRA	GLCM texture features; selected via MbPSRA	Residual U-Net	Claimed better performance than other state-of-the-art; specifics not numerically detailed
Sharifi et al. (2025)^38	Automate risk stratification of thyroid nodules using ACR-TIRADS-based DL framework.	2450 US images, 3250 nodules (1037 patients, 2018–2020)	Preprocessing, artifact removal, augmentation, Faster R-CNN ResNet-101, Xception	ACR-TIRADS features (composition, margin, echogenicity, shape, echogenic foci)	Faster R-CNN ResNet-101 (detection), Xception (classification)	Accuracy: 98.3%, AUC: 0.991, F1-score: 0.94, Sensitivity: 91%, Specificity: 97%
Al Khamsan et al. (2025)^40	Distinguish benign/malignant thyroid tumors using feature fusion network.	1137 US images (594 benign, 543 malignant)	ADF preprocessing, EfficientNetB0 + custom parallel CNN	Deep fused features from parallel CNN paths + EfficientNetB0	SoftMax classifier (fully connected layers)	Accuracy: 92.28%, Recall: 90.97%, F1-score: 92.76%, Specificity: 96%, PPV: 93%
Haider et al. (2025)^37	Develop Dual-OptNet combining ResNet and InceptionV3 with dual optimization (Adam + SGD) for thyroid disease classification.	Kaggle + Local Hospital Datasets (multiple thyroid disease types)	Modified ResNet + InceptionV3, dual optimizers (Adam and SGD), data augmentation	Multi-scale features (ResNet skip connections + Inception modules)	Dual-OptNet (CNN with dual-step optimizer)	Accuracy: 97% (average and best); improved generalization and robustness
Yang et al. (2025)^39	Enhance thyroid nodule segmentation with self-attention Swin U-Net to improve precision in ultrasound imaging.	Ultrasound images from Jinjiang Hospital and other clinical sources	Swin U-Net with multi-scale convolution, self-attention, skip connections	Multi-scale spatial features, edge-preserving attention modules	Swin U-Net (self-attention-based segmentation)	Dice Similarity Coefficient (DSC): 0.78; better than U-Net and DeepLabv3 +
Zheng et al. (2025)^41	Assess thyroid nodule malignancy using multi-subnet attention-based CNN (UMAM-NET)	Public and private datasets with annotated ultrasound images	UMAM-NET: Multi-subnet attention mechanism (shape + texture)	Texture + shape features using parallel subnet attention mechanisms	UMAM-NET classifier (malignancy grading)	Accuracy: 98.4%, F1-score: 95.4%, Recall: 93.1% (public); 94.4% (private)

Cross-sectional limitations analysis

A synthesis of general limitations among studies reviewed identifies some common problems.

Dataset and annotation

A number of these studies are grounded in small single-center datasets with low diversity. That restricts statistical power and introduces bias. Additionally, population and device dependency reduce generalizability. Annotation quality is yet another problem, as labels usually come from single raters whose histopathological verification is inconsistent.

Generalizability

There is not much external verification with few models having been tested on distinct datasets. Imbalance between malignant and benign cases lowers robustness even more. Robustness across domain shift from hospitals and instruments is mostly not studied.

Interpretability

Deep models like CNNs and transformers perform well but act as black boxes with little clinical explanation. There are not many studies involving clinician verification or showing quantifiable impact on diagnostic decision-making.

Reproducibility and reporting

Results are not always reported regularly, with irregular use of accuracy, AUC, Dice, and the like without confidence intervals often. There are not many studies that include code or data, thereby limiting reproducibility. Pipelines for preprocessing are not defined enough to facilitate replicability.

Deployment and clinical integration

Extremely high computational intensity prohibits use in resource-limited settings. Retrospective analyses dominate with very few prospective trials or clinical confirmations. Regulatory, safety, and confidentiality considerations are scarcely mentioned, hindering practical implementation.

Recommendations for future research

Future studies will need to employ multicentre datasets involving variable populations and equipment along with rigorous external verification to ensure generalizability. Improvement in the quality of annotations by consensus approaches and histopathological verification whenever feasible will reduce label noise and ensure reliability.

Systematic schemes of reporting are also vital. Researchers must always report such measures as accuracy, AUC, sensitivity, and specificity with confidence intervals so that objective comparisons between studies can take place. Interpretability tools such as attention maps or saliency analyses, validated directly with clinicians to instill confidence in the models, can also enhance transparency.

Of equal significance is encouraging reproducibility. Public availability of pre-processing pipelines, pretrained weights, and code will enable other researchers to validate findings and take the field forward as a collective. Finally, considerations for deployment cannot be avoided. Reporting model sizes, hardware requirements, and time for inference ensures that proposed systems can reasonably be implemented into practice in resource-poor environments.

Conclusion

Ultrasound is an inexpensive, real-time, and safe technology often utilized for thyroid disorder early diagnosis, which is increasingly on the rise. This article discussed how ultrasound images can be utilized in computer-aided diagnosis systems to detect and categorize thyroid problems. Deep learning already has demonstrated considerable promise for the identification, characterization, and predictive modeling of thyroid nodules. In the examined studies, deep learning algorithms particularly have exceeded traditional machine learning approaches but often are limited by limited institution-based datasets, lack of external validation, and variable reporting. These shortcomings make essential the construction of larger multicenter datasets and standardized evaluation methods to yield robust data amenable to clinical validation.

Looking ahead, the promise of AI in thyroid disease will extend well beyond image-based classification. Future potential uses should include combination of AI with electronic health records for integrated risk modeling on a holistic basis, point-of-care decision-support at the time of ultrasound evaluation, use on telemedicine platforms for increased access to specialist expertise, and use for longitudinal evaluation of thyroid nodules with insights for personalized treatment planning and follow-up strategies. These technologies will become possible with close partnerships between engineers, clinicians, and regulatory officials that will work towards developing AI technologies that are interpretable, transparent, and safe for deployment at the bedside. Lastly, while AI has already revolutionized thyroid imaging science, its broader clinical promise will depend on surmounting current weaknesses as much as embracing opportunities for scalable, dependable, and patient-focusing deployment.

CRediT authorship contribution statement

Sathya J – Formal analysis, resources, writing.

Ramkumar S – Monitoring, reviewing, editing.

All authors have read and agreed to the published version of the manuscript.

Funding

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Conflict of interest

The authors declare that they have no conflict of interest or personal relationships that could have appeared to influence the work.

Acknowledgment

The author would like to thank Dr. Ramkumar S for his guidance and comments during this work on research.