Advancing thyroid diagnosis: integrating AI-driven CAD framework with numerical data and ultrasound images

Research gap

While the aforementioned studies have made significant strides in thyroid disorder diagnosis, there is a notable gap in the integration of multiple data modalities, such as numerical patient data and ultrasound images, within a unified diagnostic framework. Existing studies often focus on either numerical data or imaging data, but not both. This limitation hinders the development of comprehensive diagnostic tools that can utilize the strengths of both data types. Additionally, there is a need for more interpretable AI models that can provide clinicians with transparent insights into the diagnostic process.

The current study aims to address these gaps by proposing a CAD framework that integrates numerical data and ultrasound images, utilizing advanced technologies such as ViTs and SHAP explainable AI to increase diagnostic accuracy and interpretability. By bridging these gaps, this study seeks to advance the field of thyroid diagnosis and provide clinicians with a more robust and transparent diagnostic tool.

Numerical feature selection via the sparrow search algorithm

Feature selection is a useful step in building effective ML models, particularly for medical diagnosis tasks such as thyroid disease classification (Balaha, Hassan & Balaha, 2025). The SSA, inspired by the foraging behavior of sparrows, is employed to optimize feature selection. The SSA mimics the exploration and exploitation strategies of sparrows to search efficiently for the most informative features (Xue & Shen, 2020).

The primary objective of feature selection via SSA is to identify the subset of input features that have the greatest impact on the prediction performance of the ML model. This approach reduces the dimensionality of the feature space, improving model interpretability and reducing computational complexity (Gad et al., 2022).

Mathematically, the SSA operates by iteratively updating the position of a population of candidate solutions (sparrows) in the search space. The position of each sparrow represents a potential solution, i.e., a subset of features selected for classification. The movement of sparrows is guided by a combination of exploration and exploitation strategies, aiming to balance the search for new regions and the exploitation of promising solutions (Gharehchopogh et al., 2023).

The movement of sparrows in SSA is governed by Eqs. (1) and (2), where $\mathrm{P}\mathrm{o}\mathrm{s}(t)$ represents the position of sparrows at iteration $t$, $\mathrm{S}\mathrm{t}\mathrm{e}\mathrm{p}\mathrm{S}\mathrm{i}\mathrm{z}\mathrm{e}(t)$ denotes the step size controlling the magnitude of movement, $\mathrm{R}\mathrm{a}\mathrm{n}\mathrm{d}$ is a random number sampled from a uniform distribution, $\lambda$ is a scaling factor, $\beta$ is a control parameter, and $\mathrm{\Gamma }$ denotes the gamma function. Algorithm 1 summarizes the major steps for feature selection via the SSA.

(1) $$\mathrm{P}\mathrm{o}\mathrm{s}(t+1)=\mathrm{P}\mathrm{o}\mathrm{s}(t)+\mathrm{S}\mathrm{t}\mathrm{e}\mathrm{p}\mathrm{S}\mathrm{i}\mathrm{z}\mathrm{e}(t)\times \mathrm{L}\mathrm{e}\mathrm{v}\mathrm{y}\mathrm{F}\mathrm{l}\mathrm{i}\mathrm{g}\mathrm{h}\mathrm{t}$$

(2) $$\mathrm{L}\mathrm{e}\mathrm{v}\mathrm{y}\mathrm{F}\mathrm{l}\mathrm{i}\mathrm{g}\mathrm{h}\mathrm{t}=\mathrm{R}\mathrm{a}\mathrm{n}\mathrm{d}\times {\left(\frac{\lambda \times \mathrm{\Gamma }(1+\beta )\times \sin (\pi \times \beta /2)}{\mathrm{\Gamma }((1+\beta )/2)\times \beta \times 2^{(\beta -1)/2}}\right)}^{1/\beta }.$$

Numerical classification via machine learning

Machine learning techniques play a pivotal role in thyroid disease classification by utilizing patient data to distinguish between benign vs. malignant cases. A diverse range of classifiers, including support vector machine (SVM), random forest (RF), AdaBoost, LightGBM (LGBM), and logistic regression (LR), are employed to achieve accurate and reliable predictions (Song, Ristenpart & Shmatikov, 2017; Balaha et al., 2024b). To increase the performance of these classifiers, various scaling techniques are applied to preprocess the input features. These include robust scaling, min–max scaling (MinMax), max absolute scaling (MaxAbs), standardization (STD), L1 normalization (L1), L2 normalization (L2), and max normalization (Max). These scaling techniques ensure that the input features are appropriately scaled and centered, thereby improving the convergence and stability of the ML algorithms (Ahsan et al., 2021).

Additionally, the TPE is employed to optimize the hyperparameters of the classifiers and scalers. It is a Bayesian-based optimization technique that efficiently searches the hyperparameter space to identify configurations that maximize the model’s performance metrics (Watanabe, 2023). The optimization process involves maximizing the expected improvement over the current best solution, as formulated in Eq. (3), where $x$ represents the hyperparameter configuration, $\mu (x)$ and $\sigma (x)$ are the mean and standard deviation of the predictive distribution, respectively, $f(x^{\ast })$ represents the best-observed value thus far, and $\zeta$ is a tunable exploration-exploitation trade-off parameter.

(3) $$EI(x)=\{\begin{array}{cc}\frac{\mu (x)-f(x^{\ast })-\zeta }{\sigma (x)},& \mathrm{i}\mathrm{f}\sigma (x)>0\\ 0,& \mathrm{o}\mathrm{t}\mathrm{h}\mathrm{e}\mathrm{r}\mathrm{w}\mathrm{i}\mathrm{s}\mathrm{e}\end{array}.$$

By iteratively evaluating the performance of different hyperparameter configurations and updating the surrogate model, TPE guides the search toward regions of the hyperparameter space that are likely to yield better results, ultimately improving the classification accuracy of the models.

Imaginary classification via vision transformers

ViTs are promising approaches for classification tasks in thyroid diagnosis via ultrasound images. Initially designed for natural image classification, ViTs have demonstrated remarkable adaptability to medical image analysis domains (Park & Kim, 2022). The core of ViTs lies in their ability to process input images through a series of self-attention mechanisms, enabling comprehensive feature extraction and representation (Raghu et al., 2021). Mathematically, the mechanism of the self-attention ( $\mathcal{A}(Q,K,V)$) can be expressed as shown in Eq. (4), where Q, K, and V correspond to the query, key, and value matrices, respectively; $d_k$ represents the dimensionality of the key vectors; and the SoftMax function calculates the weights of the attention vectors. These weights are subsequently multiplied by the value vectors to produce the attention output (Zhou et al., 2022).

(4) $$\mathcal{A}(Q,K,V)=\mathrm{S}\mathrm{o}\mathrm{f}\mathrm{t}\mathrm{M}\mathrm{a}\mathrm{x}\left(\frac{Q\times K^T}{\sqrt{d_k}}\right)\times V.$$

In the context of thyroid diagnosis, ViTs can effectively distinguish between benign and malignant thyroid nodules by processing ultrasound images through multiple layers of self-attention blocks. These blocks capture both low-level image features and high-level contextual information, enabling the model to learn hierarchical representations of the data. Each self-attention block comprises multiple attention heads that support the model to simultaneously attend to different regions of the image (Khan et al., 2022).

The feature representation obtained from the self-attention blocks is further refined through feed-forward neural network layers, facilitating classification. Mathematically, the feed-forward transformation is represented in Eq. (5), where $x$ represents the input feature representation; $W_1$, $b_1$, $W_2$, and $b_2$ denote the weights and biases of the feed-forward neural network layers, and ReLU denotes the rectified linear unit activation function (Jamil, Jalil Piran & Kwon, 2023).

(5) $$\mathrm{F}\mathrm{F}\mathrm{N}(x)=\mathrm{R}\mathrm{e}\mathrm{L}\mathrm{U}(x\times W_1+b_1)\times W_2+b_2.$$

Through extensive training on large-scale datasets, ViTs can learn discriminative features indicative of pathological characteristics associated with thyroid malignancy, such as irregular borders, microcalcifications, and increased vascularity.

eXplainable AI using SHAP: enhancing medical understanding of thyroid conditions

Explainable artificial intelligence (XAI) employing SHAP provides invaluable insights into the decision-making processes of the ML models, especially in healthcare (Arrieta et al., 2020). SHAP values offer a comprehensive view of how individual features contribute to model predictions, enhancing interpretability and trust (García & Aznarte, 2020; Aljadani et al., 2023).

Mathematically, SHAP values are represented as in Eq. (6), where ${\varphi }_i(f)$ represents the SHAP value of feature $i$ for prediction $f$, F denotes the set of all features, S represents a subset of features excluding feature $i$, and $f(S)$ and $f(S\cup \{i\})$ denote the model’s output for subsets S and S with feature $i$ included, respectively (Speith, 2022).

(6) $${\varphi }_i(f)=\sum _{S\subseteq F\backslash \{i\}}\frac{|S|!(|F|-|S|-1)!}{|F|!}\times \left[f(S\cup \{i\})-f(S)\right].$$

From a medical perspective, SHAP analysis reveals the significance of various factors in diagnosing and managing thyroid conditions. For example, in assessing thyroid disorders such as hypothyroidism or hyperthyroidism, SHAP highlights the contributions of crucial biomarkers such as TSH and thyroxine levels and patient demographics. By quantifying the impact of each feature on model predictions, SHAP enables clinicians to prioritize relevant clinical markers and tailor treatment strategies accordingly (Tjoa & Guan, 2020).

Moreover, SHAP-based explanations facilitate patient–centered care by empowering individuals to understand the rationale behind diagnostic decisions. Patients can learn how their medical history, laboratory results, and demographic characteristics influence the likelihood of having thyroid disease. This transparency fosters trust between patients and healthcare providers, ultimately leading to more collaborative decision-making processes and improved health outcomes.

Evaluation method

To assess the performance of ML models in classifying thyroid diseases, a range of evaluation metrics are utilized, such as accuracy (ACC), balanced accuracy (BAC), precision (PRC), recall (REC), specificity (SPC), F1-score, and intersection over union (IoU). Each metric offers distinct insights into different facets of the model’s classification performance (AbdulAzeem et al., 2025a). Accuracy reflects the overall correctness of the model’s predictions, while BAC addresses class imbalance by averaging the accuracy of both classes. Precision evaluates the ratio of true positive predictions among all positive predictions, and recall (or sensitivity) measures the ratio of true positives identified out of all actual positive cases (Balaha et al., 2025). Specificity assesses the model’s ability to classify negative cases correctly, and the F1-score provides a harmonic balance between precision and recall. Lastly, IoU measures the overlap between predicted and ground truth labels, making it particularly valuable for tasks like semantic segmentation (Adhikari et al., 2021).

To ensure the robustness of the performance evaluation, cross-validation is employed with both 5-fold and 10-fold schemes. Each cross-validation iteration divides the dataset into multiple subsets, with a portion reserved for training and the rest reserved for validation. This process is repeated numerous times, with different subsets used for training and validation in each iteration. The model’s generalization ability can be reliably assessed by averaging the performance metrics across multiple cross-validation folds (Balaha et al., 2024a).

To account for variability in performance due to randomness in the data and model initialization, 100 trials are conducted for each cross-validation scheme. This extensive experimentation enables the estimation of 95% confidence intervals (CIs) for each performance metric, providing a measure of the uncertainty associated with the reported results. Through this rigorous evaluation process, the reliability and generalization ability of ML models for thyroid disease classification can be accurately assessed, facilitating informed clinical decision-making and improving patient care (Kim, Hong & Yoon, 2020).

Training details of vision transformers

The ViTs were trained using the following setup:

• –

  Pretraining details: The ViT models were pretrained on large-scale datasets such as ImageNet-21k to utilize transfer learning. This pretraining step helps the model capture generic image features before fine-tuning on the specific task of thyroid ultrasound image classification.

• –

  Fine-tuning parameters: (1) Learning rate: We employed a cosine annealing learning rate scheduler with an initial learning rate of 0.001. The learning rate was gradually reduced over the course of training to ensure convergence. (2) Number of epochs: The models were fine-tuned for a total of 50 epochs. (3) Early stopping with a patience of 10 epochs was implemented to prevent overfitting and to optimize training time. (4) Batch size: A batch size of 32 was used during fine-tuning to balance computational efficiency and model performance.

• –

  Regularization techniques: (1) Weight decay: L2 regularization (weight decay) was applied with a coefficient of 0.0001 to penalize large weights and improve generalization. (2) Dropout: Dropout layers with a dropout rate of 0.1 were incorporated within the transformer blocks to further mitigate overfitting. (3) Data augmentation: Although data augmentation led to decreased performance in our experiments (as discussed in the Experiments section), standard augmentations such as random horizontal flipping and minor rotations were initially considered but later excluded from the final model training due to their negative impact on accuracy.

• –

  Optimizer: Adam optimizer was utilized for its adaptive learning rate properties, which are beneficial for training deep neural networks like ViTs.

• –

  Loss function: Cross-entropy loss was used as the objective function to optimize the classification performance of the ViT models.

• –

  Hardware and software: The models were trained on an NVIDIA GPU with 6 GB of memory, utilizing Python and relevant deep learning libraries such as PyTorch for implementation.

Computing infrastructure

The experiments for this study were designed to rigorously evaluate the performance of the proposed CAD framework for thyroid disease diagnosis. The software configuration was centered on Python, utilizing Windows 11 as the operating system and Anaconda as the distribution platform. The hardware specifications included an NVIDIA GPU with 6 GB of memory, 128 GB of RAM, and an Intel Core i7 processor, ensuring sufficient computational power for training and evaluating DL models.

Experiments on the numerical thyroid disease patient dataset

Table 1 presents the performance metrics obtained from feature selection via a 10-fold cross-validation setup. The metrics are represented as the mean values with 95% confidence intervals (CIs) derived from 100 trials on the thyroid disease patient dataset. The “Best” row highlights the performance metrics from the top-performing models employing majority voting.

Among the evaluated algorithms, AdaBoost, decision tree (DT), histogram-based gradient boosting (HGB), logistic regression (LR), and random forest (RF) emerged as the top performers. These models were optimized with specific hyperparameters, such as the regularization parameter $C=2.0288193$ for logistic regression and the learning rate $\eta =0.193226797$ for AdaBoost. The decision tree was configured with entropy as the splitting criterion, a maximum depth of 4, and the best splitter. The random forest algorithm adopted a Gini criterion, a maximum depth of 8, and 99 estimators, whereas the HGB algorithm employed a learning rate of $\eta =0.3412151542$ and a maximum depth of 10. The number of features selected varied across the models, with logistic regression utilizing nine features, AdaBoost and random forest employing 13 features each, and decision tree and HGB using 16 features. Scaling techniques further enhanced model performance, with logistic regression using L1 normalization, AdaBoost employing MinMax scaling, decision tree utilizing MaxAbs scaling, and both random forest and HGB employing robust scaling.

Table 2 presents similar results for 5-fold cross-validation. The top-performing algorithms include DTs, multilayer perceptrons (MLPs), RFs, and support vector machines (SVMs). These models were assessed on the basis of their classification accuracy and robustness across different hyperparameter configurations. The optimal hyperparameters for each model were determined through extensive experimentation, with MLP employing a ReLU activation function, 256 hidden layers, and an Adam solver; RF utilizing a Gini criterion, a maximum depth of 8, and 86 estimators; DT employing a Gini criterion, a maximum depth of 18, and the best splitter; and SVM utilizing a radial basis function kernel with a regularization parameter $C=1.59$.

Figure 2 visualizes the feature importance of the 5- and 10-fold approaches for the thyroid disease patient dataset. The features were assessed for their importance in predicting thyroid conditions. Among the features, “thyroxine”, “thyroidsurgery”, and “TSH” were identified as the most important, each with an importance score of $>80\mathrm{\%}$ for ten folds and $>65\mathrm{\%}$ for five folds. Similarly, “T4” exhibited less importance between 60% and 70% for the ten folds and between 50% and 60% for the five folds.

The results highlight the pivotal role of “thyroxine”, “thyroidsurgery”, and “TSH” from the medical perspective of thyroid conditions. Thyroxine ( $T_4$) is a hormone crucial for regulating metabolism, energy production, and growth. Its high importance in predicting thyroid conditions suggests its importance in diagnosing and managing thyroid disorders, such as hypothyroidism or hyperthyroidism. Thyroidsurgery indicates a history of surgical intervention, possibly due to thyroid nodules, tumors, or other thyroid-related pathologies. This feature underscores the clinical importance of past surgical procedures in understanding and treating thyroid disorders. TSH, a thyroid-stimulating hormone, is a key marker of thyroid function, with elevated levels indicative of hypothyroidism and suppressed levels suggestive of hyperthyroidism. Its prominence underscores its fundamental role in thyroid diagnosis and monitoring.

Experiments on the DDTI: thyroid ultrasound images dataset

Table 3 lists the performance metrics derived from the DDTI (Thyroid Ultrasound Images) dataset employing ViTs, with a particular focus on the impact of data augmentation. The models are categorized on the basis of whether augmentation was employed or not, and various evaluation metrics, including accuracy, precision, recall, specificity, and Yule’s Q coefficient, are presented.

In analyzing the outcomes, it becomes evident that models trained without data augmentation consistently outperform their augmented counterparts, particularly in the context of ultrasound imagery. Notably, the models denoted as P16-224 and P32-384 exhibit superior performance, boasting accuracy scores of 95.00% and 95.06%, respectively, alongside commendable precision, recall, specificity, and F1-scores. These results underscore the efficacy of these models in accurately classifying thyroid ultrasound images without the need for data augmentation.

However, upon introducing augmentation (✓), a significant decline in performance is observed across all the metrics. For example, the accuracy decreases to 68.64%, 74.49%, and 75.71% for the augmented models P16-224-In21K, P16-224, and P32-384, respectively. This decline raises pertinent concerns regarding the applicability and efficacy of data augmentation in the domain of ultrasound imaging.

One of the primary concerns revolves around the intricate nature of ultrasound data. Ultrasound images inherently possess unique characteristics, such as varying tissue densities, acoustic impedance, and imaging artifacts, which can hinder traditional augmentation techniques. Moreover, the clinical relevance and diagnostic accuracy of ultrasound images necessitate preserving the integrity and fidelity of the data, making it imperative to tread cautiously when applying augmentation.

Furthermore, the efficacy of augmentation techniques in enhancing model generalizability and robustness may be contingent upon the appropriateness of the augmentation strategies employed. Given the nuanced nature of ultrasound imaging, conventional augmentation techniques may inadvertently introduce distortions or artifacts that compromise the interpretability and diagnostic utility of the images. Consequently, a delicate balance exists between augmenting the data to improve model performance and preserving the inherent characteristics of ultrasound images.

Figure 3 illustrates the interpretability achieved through SHAP on DDTI: the Thyroid Ultrasound Images dataset. The visualization shows two randomly selected samples from the dataset, each accompanied by the top three predictions generated by the model. Additionally, the figure provides insights into the interpretability of these predictions, offering a transparent view of the features contributing most significantly to each prediction. This comprehensive approach enhances the understanding of the model’s decision-making process and facilitates the identification of crucial diagnostic factors in thyroid ultrasound images.

The current study outperforms the related studies by Liu et al. (2023) and Bal-Ghaoui et al. (2023) regarding diagnostic accuracy for thyroid disorders via ultrasound imaging. Liu et al. (2023) achieved an AUC of 0.95 in receiver operating characteristic (ROC) analysis with their dynamic integration model. Bal-Ghaoui et al. (2023) attained high accuracy rates of 94.84% and 85.35% for breast and thyroid datasets, respectively, with their customized CNN architecture; the current study achieved even higher accuracy rates. Specifically, models trained without data augmentation consistently demonstrated superior performance across various evaluation metrics, with accuracy scores reaching 95.06%. These results highlight the effectiveness of the suggested approach in accurately classifying thyroid ultrasound images, surpassing the performance of existing methodologies in the field.

Clinical relevance

The experiments conducted in this study hold significant clinical relevance, particularly in the context of improving diagnostic accuracy and interpretability for thyroid disorders. Thyroid diseases, including hypothyroidism, hyperthyroidism, thyroid nodules, and thyroid cancer, pose substantial challenges to healthcare systems worldwide due to their high prevalence and potential impact on patient health. Early and precise diagnosis is critical for effective treatment and the prevention of long-term complications. The integration of numerical patient data and ultrasound images into a unified CAD framework represents a major advancement in addressing these challenges. By utilizing advanced technologies such as ViTs and SHAP, the proposed framework not only enhances diagnostic accuracy but also provides transparency in the decision-making process, fostering trust among clinicians and patients.

From a clinical perspective, the use of ViTs for analyzing thyroid ultrasound images is particularly noteworthy. Ultrasound imaging is a non-invasive and widely used modality for evaluating thyroid conditions, but its interpretation can be subjective and prone to variability among radiologists. The ability of ViTs to capture global dependencies within images and extract hierarchical features enables the model to identify subtle abnormalities that may be missed by human observers. For instance, the detection of irregular borders, microcalcifications, and increased vascularity (key indicators of malignancy) is significantly improved through the use of ViTs. This capability aligns with the clinical need for tools that can assist radiologists in making more accurate and consistent diagnoses, thereby reducing the risk of false positives and false negatives.

The incorporation of SHAP into the framework further enhances its clinical utility by providing interpretable insights into the model’s predictions. Clinicians can now understand the contribution of individual features, such as thyroxine levels, TSH, and patient demographics, to the final diagnostic outcome. This level of transparency is invaluable in clinical settings, where decisions must be justified and aligned with established medical knowledge. For example, SHAP analysis revealed that “thyroxine”, “thyroidsurgery”, and “TSH” were among the most important predictors of thyroid conditions. These findings are consistent with clinical observations, as thyroxine and TSH are well-established biomarkers for thyroid function, while a history of thyroid surgery often indicates prior pathology. By quantifying the impact of these features, SHAP empowers clinicians to prioritize relevant clinical markers and tailor treatment strategies accordingly.

Another critical aspect of the experiments is the optimization of feature selection using the SSA. In clinical practice, datasets often contain a large number of features, many of which may be irrelevant or redundant. The SSA efficiently identifies the most informative features, reducing dimensionality and improving model interpretability without sacrificing performance. This approach ensures that the CAD framework focuses on clinically meaningful attributes, such as laboratory results and patient history, which are directly actionable in diagnosis and treatment planning. Additionally, the use of the TPE for hyperparameter tuning enhances the robustness of the models, ensuring that they perform optimally across diverse clinical scenarios.

Finally, the evaluation metrics used in the experiments (such as accuracy, precision, recall, specificity, and F1-score) are highly relevant to clinical decision-making. These metrics provide a comprehensive assessment of the model’s performance, addressing both the correctness of predictions and the balance between sensitivity and specificity. The high accuracy scores achieved by the models, particularly those trained without data augmentation (e.g., 95.06% for the DDTI dataset), underscore their potential to serve as reliable diagnostic aids in real-world clinical settings. Moreover, the rigorous cross-validation and confidence interval estimation ensure that the results are generalizable and robust, further enhancing their applicability in clinical practice. By bridging the gap between advanced AI techniques and clinical needs, this study paves the way for improved diagnostic outcomes and personalized treatment strategies in thyroid disease management.