Radiomics in thyroid nodule assessment

Abstract

This narrative review was conducted by searching PubMed and Google Scholar using keywords such as ‘radiomics’ AND ‘thyroid nodules’ from March 2019 to March 2024. Inclusion criteria focused on peer-reviewed studies in English on radiomics applications; 50 studies were ultimately evaluated after excluding non-relevant works. The purpose of this article is to systematize the utilization of radiomics in assessing thyroid nodules, aiming to furnish valuable theoretical insights. As an emerging interdisciplinary domain, radiomics integrates medical imaging with computer data analysis techniques, thus constituting a potent instrument for precise diagnosis and comprehensive evaluation of thyroid nodules. Through a scrutiny of the principles, methods, and applications of radiomics in thyroid nodule evaluation, we aim to elucidate its potential for enhancing diagnostic accuracy, assessing disease progression, and mitigating overdiagnosis. Additionally, this study offers a critical assessment of the inherent strengths and limitations of the employed methodology.

Clinical trial number Not applicable.

Introduction

Thyroid nodules are defined by the American Thyroid Association (ATA) [1] as discrete lesions caused by abnormal focal growth of thyroid cells, with global prevalence > 67% [2]. The incidence of malignant thyroid nodules, particularly thyroid cancer, has been increasing internationally [3]. The majority of thyroid nodules are detected through imaging techniques, which possess a significantly higher detection rate compared to clinical examination [4]. Traditional medical imaging techniques, including ultrasound (US), computed tomography (CT), magnetic resonance imaging (MRI), and nuclear medicine, are mature but limited to anatomical judgments, lacking multi-dimensional biological insights and relying on subjective interpretation [5]. Radiomics integrates imaging and computer science to mine imperceptible features [6, 7], addressing these gaps. Although prior reviews have summarized the field, they often focus on specific modalities or methods without balanced comparisons between traditional and deep learning (DL) approaches. These features offer a comprehensive, objective basis for thyroid nodule assessment. This review provides a balanced overview of principles, DL vs. traditional methods, clinical translation, and challenges.

Methods

This narrative review employed a structured literature search to identify relevant studies on radiomics in thyroid nodule assessment. Databases searched included PubMed and Google Scholar, using keywords such as ‘radiomics’ AND (‘thyroid nodules’ OR ‘thyroid cancer’ OR ‘papillary thyroid carcinoma’). The search spanned March 2019 to March 2024 to focus on recent advancements. Inclusion criteria were English-language peer-reviewed articles reporting original research on radiomics applications (traditional machine learning [ML] or DL) in thyroid nodules; exclusion criteria included case reports, reviews, non-human studies, or those not focused on radiomics. Approximately 150 articles were initially retrieved. After screening titles and abstracts, 80 were selected for full-text review, resulting in 50 studies ultimately evaluated for their methodological rigor, findings, and clinical relevance (Fig. 1). Included studies were appraised using Radiomics Quality Score (RQS) and supplemented with the QUADAS-2 tool for risk of bias assessment, with scores < 10 indicating low quality (e.g., due to retrospective bias or lack of external validation). Appraisal revealed considerable variability, with average RQS percentages around 26% (IQR 11.5%) in similar radiomics literature [8], highlighting common issues such as retrospective designs, small sample sizes, and absence of external validation; this helped us critically assess potential biases and differentiate robust multi-center studies (higher RQS) from preliminary single-center ones.

Fig. 1.

Literature review flow diagram

Relevant terminology

Within the field of radiomics, one frequently encounters terms such as artificial intelligence (AI), ML, and DL. However, numerous studies often use the terms AI and ML interchangeably, potentially leading to confusion and imprecision in communication [9]. This paper seeks to elucidate both the distinctions and the connections between these terms.

AI is a macro concept which necessitates that computers have the ability to learn and think like humans. As an important area of AI, ML explores how computers can enhance their performance autonomously without explicit programming. The aim is to devise algorithms capable of utilizing various feature combinations to predict specific outcomes [10]. Two primary categories of ML are identified: supervised learning (SL) and unsupervised learning (UL). In SL, radiomics features are labeled, learning occurs from labeled data, and labels are assigned to new unlabeled data [11]. Examples include Naive Bayes Classifier (NB), Logistic Regression (LR), Support Vector Machine (SVM), Decision Tree (DT), Random Forest (RF), and Gradient Boosting Decision Tree (GBDT). These methods are commonly employed in modeling for thyroid nodule radiomics research. UL employs data compression, downscaling, and feature filtering within clustering methodologies to categorize radiomics features [12]. In UL, input data is devoid of labels, necessitating advanced computational analysis to discern the inherent structural relationships within the data. The goal of UL is the discovery of the data’s intrinsic dimensions, components, clusters, and more [11]. Algorithms encompassed in this context include K-Means Clustering, Hierarchical Clustering, and Principal Component Analysis (PCA).

The Artificial Neural Network (ANN) is a mathematical model replicating the human brain’s neural structure [13]. It simulates learning and memory through adjusting connections between neurons, enabling tasks like classification and regression. With deepening research, the Deep Neural Network (DNN) emerged, forming deeper structures by adding hidden layers. This allows it to iteratively extract complex data features, facilitating learning of sophisticated representations. Notably, following the coinage of the term “neural network,” it became conventional to refer to earlier ML methods as “traditional ML,” paralleling the terminology of “traditional radiomics.”

DL revolutionizes ML, building on ANN. DL algorithms automate data feature learning, eliminating manual engineering. Common algorithms are Convolutional Neural Networks (CNN), Recurrent Neural Networks (RNN), Long Short-Term Memory (LSTM), Generative Adversarial Networks (GAN) and Transfer Learning. It is noteworthy that Transfer Learning is occasionally employed in thyroid nodule radiomics research, given its ability to enhance performance, reduce training duration, and minimize resource demands.

Radiomics constitutes a significant and rapidly growing subfield within the domain of AI [14]. It integrates both traditional ML and DL techniques, significantly enhancing the analysis and interpretation of medical images. In summary, ML provides a methodology for discerning patterns and models from data, focusing on algorithms and models that enable machines to learn from data and improve performance. Radiomics applies techniques such as ML to medical image processing, emphasizing the extraction of features and the analysis of diseases from medical image data (Fig. 2).

Fig. 2.

AI-ML-DL-ANN hierarchy schematic

Fundamental steps in traditional radiomics

Data acquisition

The initial task in radiomics research involves the acquisition of high-quality medical images for preliminary study data. These data may originate from a range of sources, including multiple centers, technologies, and parameters, thus reflecting the inherent diversity and complexity of radiomics. Imaging techniques recommended by the ATA guidelines include US, CT, MRI, positron emission tomography (PET), especially 18 F-fluorodeoxyglucose PET/CT (18 F-FDG PET/CT). US imaging, providing data on morphology, echotexture, and vascularity, is gaining popularity in thyroid radiomics [15], yet it faces challenges imaging deep lesions and capturing all lesion characteristics with 2D images [16]. CT imaging, favored for its high spatial resolution and standardized methodology, accurately depicts 3D morphology, tissue density, blood supply, and thyroid-structure relationships. Dual-energy CT provides iodine content data, enriching CT-based radiomics. MRI excels in soft tissue imaging, offering multi-parametric, radiation-free imaging and integrating with advanced modalities for comprehensive tumor profiling. 18 F-FDG PET/CT, recognized for distinguishing benign/malignant thyroid nodules and predicting metastases [17], accurately observes glucose metabolism and enables semi-quantitative SUV analysis (Tables 1 and 2).

Table 1.

Overview of selected radiomics studies related to thyroid nodules

Application	IMAGE	Number of patients	Segmented region (cropped image)	Feature Selection	Multi-center/Prospective/External validation (Y/N)	Model	Performance*	Feature Families (first-order/shape/texture/wavelet/deep)	Dataset Availability (public/private/on-request)	Reference
Lymph node metastasis prediction	CT (VP)	461	whole thyroid gland (3D)	Mann-Whitney U+LASSO	N/N/N	Nomogram	AUC (95% CI): 0.797 (0.690–0.902)	First-order, shape, texture	Private	[18]
Classification (benign and malignant nodes)	CT (AP)	407	The contours of the nodules (3D)	Mann-Whitney U-tests+ LASSO+ Spearman correlation coefficient	N/N/N	ML	AUC (95% CI): 0.877 (0.818 ~ 0.935) Accuracy (%): 80.49% Sensitivity (%): 81.54% Specificity (%): 79.31%	Shape, texture, wavelet	Private	[19]
Lymph node metastasis prediction	CT (NE + CE)	134	lymph nodes (3D)	Boruta algorithm	Y/N/N	ML	SVM AUC (95% CI): 0.932(0.881 ~ 0.999) Accuracy (%): 86.7% Sensitivity (%): 95.1% Specificity (%): 78.6%	First-order, texture, wavelet	Private	[20]
Classification (benign and malignant nodes)	US	508	-	-	N/N/N	DL	Accuracy: 90.20% Specificity (%): 83.3% Precision: 89.47% Recall: 1 F1-score: 94.44%	Deep	Private	[21]
Classification (benign and malignant nodes)	US	1609	Thyroid nodules (2D)	LASSO	N/N/N	radiomics score	AUC (95% CI): 0.75 (0.69 ~ 0.81)	First-order, texture	Private	[22]
Classification (benign and malignant nodes)	US	1040	The square encompassed the nodule.	MI + LDA	N/N/N	ML and DL	DL model AUC (95% CI): 0.7127 (-) Accuracy (%): 74.69% Sensitivity (%): 63.10% Specificity (%): 80.20%	Texture, deep	Private	[23]
Lymph node metastasis prediction	US	3172	-	-	Y/N/Y	TLR	Independent testing set 1 AUC: 0.93 Accuracy: 86% Specificity: 89% Sensitivity: 83% F1-score: 87%	Deep	Private	[24]
Thyroid Capsule Invasion Prediction	CT (NE + AP)	412	Thyroid nodules (3D)	ANOVA+LASSO	Y/N/Y	ML	combined-L-SVM model AUC (95% CI): 0.776 (0.625–0.904)	Shape, texture, wavelet	Private	[25]
Extrathyroidal Extension Prediction	CT (NE + VP)	199	Thyroid nodules (3D)	ANOVA+LASSO	Y/N/Y	Nomogram	AUC (95% CI): 0.797 (0.665–0.926) Accuracy: 84.8% Sensitivity: 69.2% Specificity: 90.9%	First-order, shape, texture	Private	[26]
Classification (benign and malignant nodes)	US	(631 thyroid nodules from two centers)	the rectangular area encompassing the nodule	-	Y/N/Y	DL	Xception classification model AUC (95% CI): 0.970 (0.940, 1.000)	Deep	Private	[27]
Automatic segmentation of different tissue components	US	234	-	-	N/Y/N	Nomogram	Mean Dice coefficient: 0.76 Mean true positive fraction:0.90 Mean false positive fraction:1.61 × 10 − 6	Texture, deep	Private	[28]
BRAFV600E mutation prediction	US	810	Thyroid nodules (2D)	-	Y/N/Y	DL	“Test set 1” AUC (95% CI): 0.93 (0.88–0.96) Accuracy (%): 85.1% Sensitivity (%): 81.9% Specificity (%): 88.3%	Deep	Private	[29]
Lymph node metastasis prediction	CT (NE + VP)	113	lymph nodes (3D)	-	N/Y/N	DL	Ensemble model AUC (95% CI): 0.829 (0.732–0.927) Accuracy (%): 82.3% Sensitivity (%): 83.9% Specificity (%): 76.9%	First-order, texture, deep	Private	[30]
Classification (benign and malignant nodes)	US	765	-	-	N/N/Y	ML	Accuracy (%): 82.1% Sensitivity (%): 91.9% Specificity (%): 78.5%	Texture, wavelet	Private	[31]
Classification (benign and malignant nodes)	CT (NE)	880	Thyroid nodules (3D)	-	N/N/N	DL	Ensemble model AUC (95% CI): 0.947 (0.906–0.974) Accuracy (%): 85.9% Sensitivity (%): 91.9% Specificity (%): 82.1%	Deep	Private	[32]
Lymph node metastasis prediction	US + MRI (T2WI, DWI, CE-T1WI)	133	lymph nodes (2D)	recursive feature elimination algorithm	N/N/N	Nomogram	AUC (95% CI): 0.87 (0.75–0.99)	Texture, wavelet	Private	[33]
Classification of incidental thyroid nodules	CT (NE)	139	The largest contour of the nodule (2D)	ANOVA + KW + RFE	N/N/N	ML	AUC (95% CI): 0.78 (0.60 ~ 0.93) Accuracy (%): 90.9% Sensitivity (%): 68.9% Specificity (%): 83.3%	First-order, shape	Private	[34]

NE = non-enhanced; CE = Contrast Enhancement; AP = Arterial Phase; VP = Venous Phase; MI = Mutual Information; LDA = Linear Discriminant Analysis

*The optimal performance model, validated by the validation set or external test set

Table 2.

Best-practice checklist for reviewed studies

Patient-Level Splitting (train/val/test ratios)	Data Leakage Controls	Cross-Validation Strategy	Hyperparameter Tuning Approach	External Validation (different center/device/data distribution)	Code/Model Sharing (Y/N; open-source info if available)	Evaluation Metrics	Reporting Standardization	Reference
70/15/15	Feature normalization; no patient overlap	5-fold CV	Grid search	N (single-center)	N	AUC, Accuracy, Sensitivity, Specificity	Partial (no PRISMA)	[18]
80/10/10	Random splitting; leakage checks	Nested CV	Random search	N	N	AUC, Accuracy, Sensitivity, Specificity	Partial	[19]
70/20/10	Stratified splitting	10-fold CV	Grid search	Y (multi-center)	N	AUC, Accuracy, Sensitivity, Specificity	Good (RQS used)	[20]
75/15/10	No overlap; temporal split	k-fold (k = 5)	Bayesian optimization	N	N	Accuracy, Precision, Recall, F1	Partial	[21]
80/20 (no val)	Feature scaling	None	Manual	N	N	AUC	Partial	[22]
70/15/15	Random; checks for ID overlap	5-fold CV	Grid search	N	N	AUC, Accuracy, Sensitivity, Specificity	Partial	[23]
60/20/20	Stratified; leakage prevention	Nested CV	Random search	Y (multi-center)	N	AUC, Accuracy, Sensitivity, Specificity, F1	Good	[24]
70/15/15	Random splitting	10-fold CV	Grid search	Y	N	AUC	Partial	[25]
75/25 (no val)	No overlap	k-fold (k = 5)	Bayesian	Y	N	AUC, Accuracy, Sensitivity, Specificity	Partial	[26]
80/10/10	Stratified	Nested CV	Grid search	Y	N	AUC	Good	[27]
70/20/10	Temporal split	5-fold CV	Manual	N (prospective)	N	Dice, TPF, FPF	Partial	[28]
60/20/20	Random; checks	10-fold CV	Random search	Y	N	AUC, Accuracy, Sensitivity, Specificity	Good	[29]
75/15/10	Stratified	Nested CV	Grid search	N (prospective)	N	AUC, Accuracy, Sensitivity, Specificity	Partial	[30]
80/10/10	No overlap	5-fold CV	Bayesian	Y	N	Accuracy, Sensitivity, Specificity	Good (SHAP reported)	[31]
70/15/15	Random splitting	k-fold (k = 10)	Grid search	N	N	AUC, Accuracy, Sensitivity, Specificity	Partial	[32]
75/25 (no val)	Feature normalization	None	Manual	N	N	AUC	Partial	[33]
80/20 (no val)	Stratified	5-fold CV	Random search	N	N	AUC, Accuracy, Sensitivity, Specificity	Partial	[34]

Segmentation of the region of interest

Firstly, the selection of segmentation regions sets the upper limit of the radiomics model [35], as subsequent feature data are derived from it. In most studies, the region of interest (ROI) is segmented along the edge of the lesion, while some studies avoid certain areas such as the region 1 mm away from the edge, intralesional hemorrhage, and calcification components. Additionally, some studies have used the entire thyroid gland [18, 36] as the ROI for research. Furthermore, different studies make choices between 2D and 3D (Volume of Interest, VOI). Typically, the segmentation of 2D regions of interest is more rapid. However, three-dimensional VOI can provide radiomics features with better reproducibility and stability [19].

Common segmentation techniques include automatic, semi-automatic, and manual methods. Manual segmentation is frequently regarded as the gold standard [37]. However, manual segmentation is time-consuming and subject to operator bias, potentially resulting in unstable radiomics features [38]. In contrast, automated segmentation is more sensitive, specific, and effective [39], making it ideal for large-scale data processing and rapid analysis. Additionally, several studies have utilized semi-automatic segmentation methods for delineating ROIs [40–43]. Semi-automatic segmentation maintains the speed and objectivity of automatic methods while mitigating some of the subjectivity and inefficiency associated with manual segmentation. It should be noted that preliminary findings indicate that semi-automatic segmentation offers greater stability in feature extraction than manual methods, though it does not reach the ideal levels of automatic segmentation [44].

Feature extraction

The extraction of radiomic features from the ROI (or VOI) can be facilitated by a range of software platforms, including 3D Slicer and ITK-SNAP, among others. Radiomics features are diverse and primarily include first-order statistical features, shape features, textural features, and higher-order features.

First-order statistical features

The first-order statistical features comprehensively characterize the pixel value distribution in the lesion region, serving as the basis for quantitative image analysis. Typically encompassed within these features are key indicators such as Mean, Standard Deviation, Median, Minimum, Maximum, and Entropy. Most first-order statistical features stem from histogram analysis, pivotal for its intuitive, comprehensible nature and vast applications.

Shape features

Shape features, reflecting tumor geometry and growth, are crucial in radiomics. They cover area, circumference, diameter, sphericity, and contour irregularity. In thyroid nodule radiomics, they aid benign-malignant differentiation. Wu et al. [45] revealed that shape characteristics can differentiate between benign and malignant lesions on CT scans. Giovanella et al. [46] conducted a radiomics analysis of 18 F-FDG-positive thyroid nodules and found that “shape_Sphericity” was the best predictor. Sphericity is a measure of the similarity of the target structure to a perfect sphere.

Texture features

Texture features, or second-order statistical features, quantify tumor heterogeneity [47], capturing microstructural variations, macrostructure, and intricate details. Common examples include Gray Level Co-occurrence Matrix (GLCM), Neighborhood Gray-Tone Difference Matrix (NGTDM), and Gray Level Run Length Matrix (GLRLM). In radiomics analyses, texture features objectively characterize the pathological changes within nodules. Kong et al. [48] investigated 21 radiomics features to discern the benign and malignant characteristics of thyroid nodules, of which 19 were texture features. In the study of lymph node metastasis in thyroid cancer, a radiomics investigation utilizing MRI [49] revealed that the extended echo time of the T2-weighted imaging (T2WI) sequence facilitates the capture of richer texture features, effectively bolstering the performance of the predictive model.

Higher order features

First-order, shape, and texture features stem directly from image properties, while high-order features result from complex transformations of these basics. Filters, including low-pass, high-pass, band-pass, and band-stop, play a vital role in this. High-order features excel at revealing complex, abstract information, but their extraction and application require tailored selection and optimization. A study on thyroid nodule radiomics revealed that, despite numerous new features derived from various filters, they did not significantly improve classification performance [50], likely due to feature redundancy or irrelevance in specific datasets. To address this, hybrid feature selection strategies (e.g., combining filter and embedded methods) are recommended to prioritize informative high-order features while avoiding overfitting.

Feature selection

Due to the dimensionality and complexity of medical imaging data, directly using all extracted features often leads to overfitting, high computational costs, and reduced interpretability. Therefore, carefully filtering and selecting the most relevant radiomics features is crucial. Feature selection techniques in ML encompass three main categories: Filter methods, Wrapper methods, and Embedded methods. Examples of Filter methods include Mutual Information (MI), Linear Discriminant Analysis (LDA), and K-best. Wrapper Methods focus on evaluating various feature subsets’ effectiveness on a specific ML model’s performance. This involves directly integrating the feature selection process into the model [51]. After evaluating these models’ performances, the optimal feature subset is selected. Embedded methods seamlessly integrate feature selection into the model training process, enhancing efficiency by considering the interplay between features and target variables right from the learning phase. Unlike wrapper methods, which rely on evaluating multiple subsets of features through explicit testing, embedded methods are inherently more efficient. A prime example of an embedded method is Lasso regression. This technique employs L1 regularization, which not only helps in controlling overfitting but also effectively reduces the complexity of the model. In radiomics research, different feature selection methods are often combined, including statistical methods, to more accurately identify key features.

Model construction

The core issue in model construction lies in the rational selection of an appropriate ML classifier, considering factors such as sample size, class imbalance, and validation strategies to mitigate overfitting and ensure generalizability. NB is a probabilistic graphical model [52] celebrated for its computational efficiency and straightforwardness in implementation, though it assumes feature independence, which may limit performance in complex, correlated radiomics data. DT utilize a hierarchical tree structure to facilitate both feature selection and classification, but are prone to overfitting in small samples [53]. RF enhances classification accuracy through an ensemble approach that involves constructing multiple decision trees and aggregating their outputs, offering robustness to class imbalance but requiring larger computational resources. SVMs are adept at managing small sample sizes, and excel in nonlinear and high-dimensional problem spaces, though performance drops without proper kernel selection or in imbalanced datasets [54]. The key challenge with SVM is the identification and optimization of a hyperplane that maximizes the margin between different classes, thereby reducing classification errors. In addition to SVM and RF, LR represents yet another conventional feature-based SL classifier, valued for interpretability but sensitive to multicollinearity in high-dimensional radiomics features. It leverages logistic functions to transform linear regression outputs into probabilities, making it particularly suitable for binary classification tasks.

The intricate traditional radiomics research process extends beyond the fundamental stages mentioned earlier, with methodological variations (e.g., imaging modality, segmentation strategy) influencing reported AUCs, as evidenced in Table 1 where CT-based models often outperform US in multi-center validations due to standardization. It comprehensively incorporates, but is not confined to, the preliminary stages of image and data preprocessing, the pivotal phase of model training and optimization, and the essential evaluation of target efficacy. Therefore, researchers are required to exercise discretion in selecting methodologies and tools, taking into account the specific context and requirements of their research. Furthermore, as documented in [20], some scholars adopt a multifaceted approach by developing multiple learning models within a single research project and subsequently identifying the optimal one. This diversified strategy enables a thorough comprehension of the strengths and limitations of each method across diverse scenarios. It is noteworthy that radiomics models exhibit diverse forms, encompassing not only ML-based radiomics models but also the development of radiomics score models or nomogram models. In practical applications, radiomics scores are often incorporated into nomograms as variables or labels (radiomics labels), alongside clinical information, to construct so-called radiomics-clinical comprehensive models. This approach has the potential to enhance predictive accuracy while personalizing the diagnostic and therapeutic process.

Deep learning model construction

Deep Learning-based Radiomics (DLR) significantly refines the methodology of handling medical imaging data. Unlike traditional methods that often require manual segmentation to identify regions of interest in the images, DLR leverages DL models that proficiently process raw image data, though this automation can introduce biases if training data lacks diversity across imaging modalities or patient demographics. This capability not only streamlines the initial steps but also integrates part of the preprocessing within the model’s training phase, enhancing efficiency but raising infrastructure demands in resource-limited settings. As a result, DL models dynamically adapt to variations in incoming data, subtly enhancing the processing workflow, yet generalizability remains limited without external validation, contributing to slow clinical adoption. Furthermore, the model is equipped with dynamic sensitivity adjustments to adapt to fluctuations in image brightness and contrast, thereby handling diverse imaging environments effectively. This adaptive processing enables the model to autonomously detect and isolate pivotal features within images, significantly improving the precision of tasks related to image recognition and analysis, despite regulatory challenges in validating ‘black-box’ models for routine practice.

The performance of DL models is associated with the intricate design of network structures, posing a pivotal challenge within the realm of DLR [55]. In thyroid radiomics research, CNNs are favoured for their excellent feature extraction capabilities, but their complexity often hinders real-world deployment without simplified architectures. CNNs are mainly composed of input, hidden, and output layers. The hidden layer, which constitutes the core of the network structure, is typically designed as a multilayer system. The convolutional layer utilizes convolutional operations to capture local features, while the pooling layer reduces data dimensions, streamlines computations, and isolates more representative features. The typical design involves alternating between convolutional layers and pooling layers. Furthermore, after the convolutional layer, an activation function (such as ReLU, Sigmoid, or Tanh) is typically applied to introduce nonlinearity into the network. The fully connected layer plays a crucial role in integrating features extracted from previous layers to output the final result, with each neuron connected to all neurons in the preceding layer. Research has shown that DL can enhance the precision of classification or prediction tasks by building multiple hidden layer ML models and training them with extensive datasets [56]. It is worth mentioning that another study [21] has demonstrated that the complexity of the network structure is not necessarily proportional to the improved performance of the model. That study compared network structures with three, four, and five layers, finding that the four-layer structure, consisting of two hidden layers, exhibited the most favourable performance (Fig. 3).

Fig. 3.

Flowchart of radiomics

From a clinical perspective, DL models face barriers in real-world integration, including high computational and infrastructure requirements (e.g., GPUs for training), limited generalizability across diverse populations and scanners, and regulatory considerations (e.g., FDA approval). Despite promising diagnostic performance (AUCs > 0.90 in controlled studies), limited adoption in routine practice stems from these issues, alongside the black-box nature reducing physician trust and lack of large-scale prospective validations.

Application

Utilizing PubMed and Google Scholar, we conducted a literature search and reviewed studies related to thyroid nodule radiomics research over a five-year period (March 2019 - March 2024). The research in the field of thyroid nodule assessment has primarily concentrated on applications such as differentiating between benign and malignant nodules, predicting lymph node metastasis (LNM), analyzing extrathyroidal extension (ETE), exploring BRAF gene mutations, and conducting research related to nodule components.

Differentiation of benign and malignant thyroid nodules

In the clinical evaluation of thyroid nodules, the differentiation between benign and malignant is of crucial importance. Despite the low incidence of malignancy (10 to 15% [57]), factors like physician conservatism and patient anxiety may lead to overdiagnosis, resulting in unnecessary fine needle aspiration biopsy (FNAB) and surgeries [58]. Integrating radiomics scores with guidelines (e.g., ATA or ACR) can improve PPV and reduce FNAB [22], though benefits are more pronounced in large cohorts. In specialized nodule types like cystic, ML models (e.g., KNN with AUC 0.909 [59]) show promise, but contradictions arise in smaller datasets (n < 200), where specificity often drops (e.g., 36.1% in adolescent studies [60]), likely due to class imbalance and overfitting. Imaging modality influences results; US-based models show higher variability from operator dependence, while CT offers better reproducibility but radiation exposure limits. Collectively, DL models frequently show improved performance over traditional ML in accuracy and AUC, particularly in large, well-validated datasets due to automated feature learning; however, in smaller or specialized cohorts (n < 200), such as adolescents, specificity remains low (e.g., 36.1% [60]), likely from data limitations, class imbalance, and overfitting risks. Segmentation strategy also plays a role—manual methods introduce bias, reducing generalizability compared to automatic approaches. This highlights an important challenge: high sensitivity at the cost of specificity in niche populations, an unresolved issue tied to limited annotated data for training and lack of external validation. Targeted radiomics studies [23, 40, 46] address this, with Luca Giovanella et al. combining radiomics with PET metrics for a multiparametric model predicting cancer prevalence (5%, 37%, and 79% for 0–2 positive biomarkers). Such multimodal approaches show promise but reveal unresolved issues in standardization, as varying modalities lead to inconsistent AUCs across studies. Yongfeng Wang et al. [61] compared traditional radiomics and DL models on 3120 ultrasound nodule images, with the modified VGG16 DL model outperforming traditional models (accuracy 74.69%, AUC 0.7127). Similarly, Wenjun Li et al. [56] developed a Unet + CNN-F model achieving 85.92% accuracy, and Ma et al. [62] used image enhancement with DL for 90.20% accuracy.

Prediction of lymph node metastasis

In the preoperative assessment of malignant thyroid nodules, CT was often regarded as the primary means of evaluating LNM. Concurrently, the utilization of CT images for radiomics studies has also garnered significant attention among researchers. Dan Kong et al. [63] employed CT radiomics to assess LNM in malignant thyroid nodules, constructing four models, with the best nomogram achieving AUCs of 0.98 (training) and 0.95 (test). This high performance contrasts with lower AUCs in other studies, likely due to phase-specific feature extraction (arterial/venous). Siyuan Lu et al. [39] built a clinical-radiomics nomogram using dual-energy CT venous images, with AUC of 0.85 (training) and 0.797 (test). Variations in AUC across studies (0.795–0.98) stem from imaging phases and clinical integration, with nomograms consistently outperforming single-modality models by leveraging multimodal data. Yan Zhou et al. [64] used dual-energy CT iodograms for nomograms predicting LNM (AUC 0.807) and CLNM (AUC 0.795) in validation. This slight drop from training highlights overfitting risks, a common unresolved issue in radiomics requiring larger external validations. Yang Gao et al. [65] developed a multimodal model (clinical, US, traditional radiomics, DL) for LNM in 613 PTC patients. Similarly, Inhuman Bu et al. [66] used a transfer learning radiomics (TLR) model achieving AUC 0.93 and 84% accuracy. These studies show DL-integrated models excel in multi-center data, but contradictions in specificity arise from data heterogeneity, underscoring the need for standardized preprocessing.

Prediction of extrathyroidal extension

ETE signifies that the tumor penetrates the thyroid gland’s capsule and invades adjacent tissues or organs, typically indicating significant aggressiveness in thyroid nodules. This condition exacerbates both the complexity and the risk of the disease. Radiomics models have shown potential in predicting ETE, with studies using CT and MRI data. For instance, one study developed a radiomics nomogram based on CT features, achieving AUCs of 0.85–0.90 in validation sets, influenced by sample sizes > 400 and external validation [67]. However, smaller, single-center studies (n < 200) report inflated AUCs due to retrospective bias and lack of class balancing, leading to poor generalizability. Manual segmentation in these studies often introduces variability, while automatic methods improve reproducibility but require computational resources. Compared to traditional ML, DL approaches show slight superiority in multi-modal data but underperform in small datasets due to overfitting.

Calcification and colloid classification

Calcification plays a crucial role in the evaluation of thyroid nodules, serving as key indicators for determining whether they are benign or malignant. Chen Chen et al. [24] optimized a YoloV5 DL model for discriminating calcifications and colloid cysts, achieving average sensitivity 78.41%, specificity 91.36%, accuracy 77.81% in one group, and lower in another (sensitivity 70.17%). This variability underscores contradictions in performance for small echogenic foci, likely due to image resolution and annotation quality. In a multi-center study [25], the Xception CNN achieved AUC 0.970, surpassing radiologists for calcified nodules. DL models like this consistently outperform traditional methods in classification tasks, but require large datasets to mitigate overfitting. Viksit Kumar et al. [68] developed an MPCNN DL model for segmenting nodules, cysts, and glands, with Dice 0.76, comparable to traditional DRLS but fully automated. This automation resolves manual bias issues in traditional segmentation, representing a key trend toward efficiency, though computational demands remain an unresolved challenge.

Exploration of BRAF gene mutation

Radiomics has been applied to predict BRAF V600E mutations in papillary thyroid carcinoma, a key prognostic marker. Studies using US and MRI features have developed models with AUCs ranging from 0.80–0.93 [26, 69]. High performance in larger cohorts (n > 500) contrasts with lower specificity in smaller samples, often due to selection bias and retrospective designs. Texture and high-order features from MRI provide richer information than US, but modality differences lead to inconsistencies. DL models, such as CNN-based, outperform traditional ML in feature automation but require large datasets to avoid overfitting; external validation is rare, limiting clinical trust.

Research related to nodule components

Radiomics aids in analyzing thyroid nodule components, such as calcification and colloid, for better risk stratification. Studies using US and CT have segmented components with DL, achieving Dice coefficients of 0.76–0.90 [25, 27, 69]. Automatic segmentation strategies excel in large datasets but face challenges in class-imbalanced small samples (n < 300), leading to false positives. Multi-center validations improve generalizability, but retrospective biases inflate metrics. Integrating with clinical data enhances cost-effectiveness by reducing unnecessary interventions, though prospective data is needed.

Discussion

Limitations

Thyroid nodule radiomics holds immense potential, yet deepening studies reveal challenges. By analyzing current research, we critically assess advantages, limitations, and outlook, particularly for clinical pain points and future prospective validations. Key limitations include data heterogeneity (e.g., drops in external AUCs due to scanner variations), class imbalance reducing specificity, and retrospective designs introducing bias. Small sample sizes (n < 200) often yield unrealistically high AUCs (> 0.90), warranting skepticism via tools like RQS or QUADAS-2. Solutions include standardization protocols and multi-center prospective studies.

Traditional radiomics and deep learning

Therefore, we must maintain a balanced and rational perspective towards traditional radiomics and DL techniques. We should harness the strengths of DL in feature extraction and classification, but also be mindful of its limitations, including computational requirements (e.g., high GPU needs for training) and infrastructure barriers in resource-limited settings. The selection of the most appropriate method should align with the specific challenges at hand, considering generalizability (e.g., external validation gaps) and regulatory considerations (e.g., FDA approval for medical devices). Despite high performance, limited clinical adoption arises from these factors plus lack of prospective, multi-center trials. Notably, some studies have successfully integrated the merits of both DL and traditional radiomics [65], providing valuable insights for future research. Quantitative synthesis from Table 1 shows mixed results; e.g., SVM AUC 0.932 [20] outperforms DL 0.713 [61] in small samples, with DL mean AUC 0.886 (n = 6) vs. traditional 0.822 (n = 8; t=-1.40, p = 0.199, not significant), indicating that DL performance advantages are context-dependent rather than universal.

Radiomics scores and nomograms

In our review of the relevant literature, we found that when assessing the aggressiveness of thyroid nodules using radiomics methods, researchers preferred to adopt the radiomics score model and the nomogram model. Aggressiveness, as a key characteristic of thyroid cancer, involves LNM, ETE, and other aspects, which have a direct impact on the progression rate of the disease and patient prognosis. Therefore, a comprehensive and personalised assessment of aggressiveness is essential for the development of precise treatment plans. The construction of a nomogram model containing radiomics labels (radscore) is a personalised and efficient assessment method. Such models take into account not only the radiomic features of the tumour, but also clinical factors such as age, gender and hormone levels. By integrating these diverse information sources, the nomogram model offers a more precise and holistic prediction, aiding doctors in swiftly assessing the impact of various factors on disease aggressiveness and their interactions. This enhances prediction accuracy, thereby providing a multi-dimensional basis for personalised management of thyroid nodules.

Application prospects

It is anticipated that the application of radiomics in the assessment of thyroid nodules will continue to be promoted and deepened in the future, as a result of the continuous innovation of technology and the continuous optimisation of algorithms.

The multicentre, prospective, external validation of radiomics models is not only a test of model reproducibility and robustness [28], but also the basis for promoting the application of radiomics models. Multi-centre studies involve different environments and equipment, as well as different operating procedures and patient populations, all of which lead to a diversity of data characteristics. Prospective study is an important paradigm in scientific research, which is a method in which the researcher predetermines the observational indicators and collects data according to the purpose of the study. This method helps researchers to collect and analyse data in a more targeted way, and makes the data structure more complete and reasonable. In contrast, prospective studies frequently employ multi-centre and large-sample designs, which facilitate further enhancement of model generalisation capabilities. While some researchers have conducted prospective studies or constructed prospective datasets [27, 29, 70, 71], they remain a minority within the relevant studies that we have reviewed.

In the field of radiomics, enhancing interpretability remains a critical challenge. To address this challenge, numerous scholars have proposed a wide range of methodologies and measures [72, 73], primarily focusing on visualization (e.g., Grad-CAM heatmaps), semantic interpretability, and local interpretability (e.g., SHAP). These have been applied in thyroid studies to improve clinical trust. Visualization techniques reveal the internal relationships and interactions within models through graphs, animations, and interactive interfaces. Semantic interpretability techniques convert model predictions into easily comprehensible semantic information, enhancing model transparency and physician trust. Local interpretability techniques, akin to a “translator” for each prediction of a ML model, provide detailed explanations of how each feature contributes to the model’s prediction. SHapley Additive exPlanations (SHAP) algorithm stands as a representative of these techniques. As a game theory-based method, SHAP estimates the contribution of each feature to a model’s prediction, facilitating the interpretation of the output of any ML model. In a radiomics study on thyroid nodules [74], the SHAP algorithm was employed to explain a ML model named Thy-Wise, enabling clinicians to better understand the diagnostic reasoning behind the model’s predictions, thereby improving the interpretability and credibility of the model’s results in clinical decision-making.

In addition, the integration of multimodal imaging information, the combination of different types of omics, and the intelligent development aided by AI and big data are expected to further elucidate the biological characteristics of nodules, discover novel therapeutic targets, and enable efficient, automated, precise, and in-depth evaluations. Emerging multimodal LLMs [75, 76] offer advantages like NLP interfaces and multi-data integration over CNNs, but face limitations in data needs, complexity, hallucination risks, and regulatory challenges. These undoubtedly constitute significant developmental directions for the research on thyroid nodule radiomics.

Clinical translation barriers

Clinical adoption barriers include computational needs for real-time analysis, physician acceptance due to black-box issues, and regulatory approval. In resource-limited settings, cloud-based deployments could help, but multi-center studies demonstrate substantial savings from reduced FNAB rates, pending outcome data. Radiomics can reduce unnecessary FNAB rates by 20–60% (e.g., from 68.5% to 9.1% in multi-center studies [30], or 35.3% to 14.5% with dual-modal US [77]), enhance risk stratification for personalized management, and improve cost-effectiveness by minimizing surgeries, though economic analyses are needed. Integration into workflows requires handling algorithmic failures via hybrid AI-human decisions, with prospective trials essential for demonstrating tangible outcomes like improved patient survival or reduced healthcare costs.

Conclusion

Radiomics represents an advanced and rapidly evolving interdisciplinary technique that can be applied in the comprehensive assessment of thyroid nodules to provide decision support for physicians. Its future application still has great potential and deserves continuous exploration and discovery.

Abbreviations

AI

Artificial Intelligence

ML

Machine Learning

DL

Deep Learning

ANN

Artificial Neural Network

SL

Supervised Learning

UL

Unsupervised Learning

NB

Naive Bayes Classifier

LR

Logistic Regression

SVM

Support Vector Machine

DT

Decision Tree

RF

Random Forest

GBDT

Gradient Boosting Decision Tree

PCA

Principal Component Analysis

DNN

Deep Neural Network

CNN

Convolutional Neural Network

RNN

Recurrent Neural Network

LSTM

Long Short-Term Memory

GAN

Generative Adversarial Network

US

Ultrasound

CT

Computed Tomography

MRI

Magnetic Resonance Imaging

PET

Positron Emission Tomography

18F-FDG PET/CT

18F-Fluorodeoxyglucose PET/CT

ROI

Region of Interest

VOI

Volume of Interest

GLCM

Gray Level Co-occurrence Matrix

NGTDM

Neighborhood Gray-Tone Difference Matrix

GLRLM

Gray Level Run Length Matrix

T2WI

T2-Weighted Imaging

MI

Mutual Information

LDA

Linear Discriminant Analysis

DLR

Deep Learning-based Radiomics

ReLU

Rectified Linear Unit

ATA

American Thyroid Association

ACR

American College of Radiology

PPV

Positive Predictive Value

FNAB

Fine Needle Aspiration Biopsy

KNN

K-Nearest Neighbors

TI-RADS

Thyroid Imaging Reporting and Data System

LNM

Lymph Node Metastasis

CLNM

Central Lymph Node Metastasis

PTC

Papillary Thyroid Carcinoma

TLR

Transfer Learning Radiomics

ETE

Extrathyroidal Extension

BRAF

B-Raf Proto-Oncogene

XAI

Explainable Artificial Intelligence

Grad-CAM

Gradient-weighted Class Activation Mapping

SHAP

SHapley Additive exPlanations

LLM

Large Language Model

NLP

Natural Language Processing

RQS

Radiomics Quality Score

QUADAS-2

Quality Assessment of Diagnostic Accuracy Studies-2

CV

Cross-Validation

TPF

True Positive Fraction

FPF

False Positive Fraction

Author contributions

W.S. and C.G. wrote the main manuscript. X.Z. and H.L. prepared the data collection. W.L. and J.Z. prepared tables. J.J. and J.Z. analyse and interpret of results. All authors reviewed the results and approved the final version of the manuscript. All authors would be informed each step of manuscript processing including submission, revision, revision reminder, etc.

Funding

The work was not funded by any funding.

Data availability

The datasets used and analysed during the current study available from the corresponding author on reasonable request.

Declarations

Ethical approval

Not applicable, as this is a narrative review of published literature with no primary data collection or involvement of human participants.

Consent to participate

Not applicable.

Consent for publication

Not applicable.

Declaration of generative AI in scientific writing

The authors did not use generative AI or AI-assisted technologies in the development of this manuscript.

Competing interests

The authors declare no competing interests.