A Comparative Study on Thyroid Nodule Classification using Transfer Learning Methods

1.1. Ultrasound Image

Ultrasound characterization of thyroid nodules pri- marily focuses on imaging features such as size, shape, echogenicity, and the presence of calcifications to determine the likelihood of malignancy. On the other hand, the study of gland function and conditions like hypothyroidism or hyperparathyroidism involves assessing hormone levels, glandular function, and associated symptoms to diagnose and manage endocrine disorders. Separating these areas of study allows for a more focused and in-depth analysis of each aspect of thyroid disease, potentially leading to more nuanced and accurate diagnostic and management strategies. Additionally, it enables researchers and clinicians to tailor interventions and treatments specifically to the needs of patients based on their presenting symptoms and diagnostic findings.

1.2. Thyroid and CAD(Computer Aided Diagnostics) System

Thyroid nodules, which afflict 20% to 69% of the general population, are among the most common nodular lesions. Thyroid cancer, one of the most prevalent types of cancer, is known to have increased by about 240% over the previous 30 years [1]. Ultrasonography is a popular and widely used examination used to diagnose the thyroid gland and forms the basis of fine needle aspiration biopsy (FNAB) and subsequent treatment. Several recommen- dations have recently been made to assist radiologists in evaluating thyroid nodules based on ultrasound. However, since ultrasound causes echogenic problems and adverse effects, ultrasound detection of thyroid nodules still depends only on the extensive training and expertise of specialists [2].

Using the proposed multi-channel paradigm, the possibility of different kernel sizes is explored and compared to single-channel CNNs. Using CT modalities in the field of thyroid disease detection, Xinyu Zhang (2022) assessed the effect of nucleus size selection on CNN performance. This study also suggests the most reliable combination of kernel sizes for CNN mode [3].

Thyroid nodules, abnormal growths in the thyroid gland, are a common occurrence in medical imaging studies, necessitating accurate identification for timely intervention [4]. Traditional methods of diagnosis involve manual interpretation of imaging data, a process susceptible to subjectivity and variability. With the rapid advancements in artificial intelligence, particularly deep learning and machine learning, there is a growing interest in leveraging these technologies to enhance the accuracy and efficiency of thyroid nodule detection [5].

Deep learning, a subset of machine learning, has demonstrated exceptional capabilities in image analysis tasks. Convolutional Neural Networks (CNNs), a pro- minent deep learning architecture, have proven successful in various medical imaging applications [6]. In the context of thyroid nodule detection, CNNs can be employed to automatically learn and extract relevant features from medical images, aiding in the differentiation of nodules [7]. The ability of deep learning models to discern complex patterns and hierarchical representations makes them well-suited for this intricate diagnostic task [8].

Few studies show that optimization techniques are crucial for refining and improving the performance of thyroid nodule detection models [9]. Hyperparameter tuning, regularization methods, and gradient descent optimization algorithms are employed to fine-tune the model parameters, ensuring optimal generalization to new data. Additionally, transfer learning, a technique where pre-trained models on large datasets are adapted to the specific task of nodule detection, enhance the efficiency of model training, particularly in scenarios with limited labeled medical data [10].

Wenkai Yang employed semi-supervised classification algorithms to generate data, addressing the reliance of deep learning models on lesion labeling and overcoming constraints posed by small datasets in medical imaging [11]. This approach facilitated improved nodule splitting and categorization. Yang also introduced a Multitask Cascade Deep Learning Model (MCDLM) that leverages multimodal ultrasound data for the automatic classi- fication of thyroid tumors. Notably, the model considers the diverse perspectives of radiologists. The Dual Knowledge (DK) framework incorporates insights from radiologists' comments and ultrasonic properties of nodules (UF), encompassing B-mode characteristics defined by TI-RADS and hardness information from elastography images. The initial nodule segmentation was performed using a U-net constructed with a pre-trained VGG13 model. Ruiguo Yu employed ultrasonic imaging is a topic of many research studies involving thyroid nodules. Thyroid nodules, which can be benign or malignant, are a prevalent ailment in coastal areas [12].

Each of them offers different therapy options and affects patients to various degrees. Although ultrasonography can be used to identify nodule features, its fuzziness and a variety of angles, as seen in, place extra demands on medical practitioners and suggest a method for CAD ultrasound pictures that is unsupervised. Lee proposed that a deep convolutional neural network (CNN) has significantly improved the efficiency and diagnostic accuracy of interpreting medical images in recent years [13]. Many studies describe a computer-aided diagnostic (CAD) model that uses computed tomography (CT) and ultrasound images to diagnose both normal and pathological thyroid conditions.

To semantically distinguish thyroid nodules in nodules, cystic components, and normal thyroid in B-mode images, Viksit Kumar came up with a breakthrough multiplex CNN [14]. The program allows the user to recognize and segment the anatomy of the thyroid in real time. The method is used to create thyroid maps, identify thyroid nodules, segment them, and estimate their volume and size. The effectiveness of the proposed method is evaluated against manual segmentation masks and traditional seeding methods.

Yinghui Lu showed that convolutional neural networks diagnose thyroid nodules more accurately than traditional artificial neural networks. However, there is still a need for enhancement in thyroid nodule placement or diagnosis accuracy [15, 16]. There is a lack of a target detection technique for deep learning networks that can locate and identify nodules in real-time. Hence, numerous studies endeavor to create a highly effective object detection model capable of discerning between benign and malignant thyroid nodules. Here are the main highlights of the proposed work:

• Comprehensive Evaluation of Transfer Learning Models: The study evaluates a range of transfer learning models, including DenseNet169, ResNet101, and various EfficientNet variants, for classifying thyroid nodules. DenseNet169 achieved the highest accuracy of 95.96%, demonstrating the effectiveness of advanced deep-learning techniques in medical image classification.
• Hyperparameter Tuning and Optimization: An exhaustive grid search strategy was employed for hyperparameter tuning, optimizing parameters such as learning rate, batch size, optimizer type, and dropout rate. This rigorous approach ensured the models were finely tuned to achieve the best performance.
• Balanced Performance Metrics: In addition to accuracy, the study monitored precision, recall, and F1 score to ensure balanced performance across different metrics. This comprehensive evaluation provided a robust assessment of each model's capabilities in distinguishing between malignant and benign thyroid nodules.
• Utilization of a Diverse Dataset: The research utilized a dataset comprising 7,893 images from four different sources (PERSIAN, DDTI, ACR TIRADS, and AUITD), ensuring a diverse and representative sample of thyroid images. This diversity enhances the generalizability and robustness of the findings.
• Implications for Improved Diagnostic Accuracy: The study highlights the potential for integrating advanced AI techniques in medical diagnostics. By outperforming traditional diagnostic approaches reliant on the human eye, these models offer a more accurate, consistent, and unbiased method for detecting thyroid diseases, potentially leading to better patient outcomes.

2.1. Expanding and Diversifying the Dataset

This study utilized a comprehensive dataset of 7,893 images from four different sources (PERSIAN, DDTI, ACR TIRADS, and AUITD), ensuring a diverse representation of thyroid images. By addressing the limitations of dataset size and diversity, the study enhances the generalizability and robustness of the model’s performance across different demographic and clinical settings.

2.2. Comprehensive Model Comparison

A wide range of transfer learning models were evaluated, including DenseNet169, ResNet101, and various EfficientNet variants. This approach provides a thorough comparison, identifying the most effective models for thyroid nodule classification and filling the gap in narrow model focus.

2.3. Rigorous Hyperparameter Tuning

An exhaustive grid search strategy was employed to systematically tune hyperparameters such as learning rate, batch size, optimizer type, and dropout rate. This rigorous optimization process ensures that the models are performing at their best, addressing the gap in insufficient hyperparameter optimization.

2.4. Balanced Evaluation Metrics

In addition to accuracy, the study monitored precision, recall, and F1 score to provide a balanced assessment of model performance. Ensuring balanced performance across different metrics helps in creating models that are not only accurate but also reliable and effective in real-world diagnostic scenarios.

3.1. Threshold

The simplest technique for segmenting images is thresholding. It can be used to create binary images from grayscale images [25]. Thresholding is a segmentation technique that creates a binary image by dividing a given grayscale image into two regions based on threshold values. A binary image is one whose pixels have only two values, 0 and 1, and so uses only one bit to record pixel intensity. So, in the output image, pixels with intensity levels more than the specified threshold will be classified as white or 1, and those with lower values will be treated as black or 0. Abbreviations and Acronyms

3.2. Noise Removal

Noise often spoils pictures. For example, noise may occur while taking and sending the photo. The role of noise removal in image processing is important. Image processing performance is often affected by noise reduction. In color image processing, a number of strategies for noise removal are well established [26]. The type of noise that affects the image determines the nature of the noise reduction problem. According to the widely recognized theory, Johnson-Nyquist noise (thermal noise), which includes noise from capacitor reset noise, is the main cause of amplifier noise. This noise is pixel-specific, additive, Gaussian, and signal-intensity dependent (“kTC noise”). Here are a few illustrations of the various types of visual noise:

1. Salt and pepper noise.

2. Shaped noise.

3. Rayleigh noise.

4. Gaussian noise.

5. Exponential noise.

6. Erlang (gamma) noise.

3.2.1. Gaussian Noise

According to the widely recognized theory, Johnson-Nyquist noise (thermal noise), which includes noise from capacitor reset noise, is the main cause of amplifier noise [27]. This noise is additive, Gaussian, pixel-specific, and signal-intensity dependent (“kTC noise”). The result of random oscillations in the signal is white noise. There may be greater noise when the blue channel of color cameras is amplified more than the green or red channels. The primary cause of the consistent noise level in the picture's black parts is amplifier noise, which contributes to the noise of an image sensor. Due to these effects, each pixel in the image will (usually) experience a small amount of change from its initial value. A histogram displays the typical distribution of noise and plots the degree of distortion of a pixel value against its frequency (Eq 1).

(1)

3.2.3. Rayleigh Noise

For a variable with a positive value, the Rayleigh distribution is a regular distribution. This often occurs when the amplitude and direction components of a vector are connected (Eq 2).

(2)

3.2.4. Exponential Noise

The probability distribution, known as the exponential distribution, is used to represent how quickly events occur in a Poisson process. It is an instance of the gamma distribution in particular (Eq 3).

(3)

3.2.5. Erlang (gamma) Noise

Theoretically, random noise can “represent” a mathematical function. A model is precisely what it sounds like it is. Based on probability, we may describe the “distribution” of the noise. It is considerably simpler to design filters to reduce noise once it has been quantified. We'll only review the various probability density functions (PDFs) in this post and get acquainted with six distinct noise models (Eq 4).

(4)

3.3. Normalization

Normalization in machine learning generally refers to the process of transforming input data into a standard or classification model to improve the performance and stability of machine learning models [15]. Normalization can help improve the convergence speed of training algorithms, reduce the impact of outliers in the data, and prevent numerical instability in the models. Natural language processing, recommender systems, image and audio recognition, and other machine learning applications frequently use normalization.

4.1. Dataset and Image Pre-processing Phases

The thyroid ultrasonography (US) dataset presented in this study is helpful. Two thousand four hundred fifty thyroid US images were gathered between 2018 and 2020 as part of the Prospective Epidemiology Research Study in Mashhad, Iran (PERSIAN), a significant national cohort study [28]. The characteristics of ACR TIRADS, developed by skilled medical professionals, are provided together with the ROI of thyroid nodules on these US images in XML format. ACR-TIRADS classifies the images in the collection into five classes (Tirads1-Tirads5). Here, we have collected for PERSIAN (2450 thyroid images), DDTI (480 images), and AUITD (3538 thyroid images). A graph within a graph is an “inset”, not an “insert”. The word alternatively is preferred to the word “alternately” (unless you really mean something that alternates). Table 1 shows the number of images in each dataset used for thyroid classification. The datasets used (PERSIAN, DDTI, ACR TIRADS, and AUITD) are curated from reputable sources, ensuring high-quality images and accurate annotations. High-quality data can often compensate for smaller datasets by providing clear and precise information for model training. Each dataset includes a representative sample of the specific medical conditions or imaging modalities under study. This targeted approach helps in building a robust model for the intended application, even with a limited number of images. To mitigate the issue of limited data, we have employed advanced data augmentation techniques. These techniques artificially increase the dataset size by generating modified versions of existing images through transformations such as rotations, scaling, flipping, and intensity variations. This helps in improving the model's robustness and generalization ability.

4.2. Workflow of the Comparative Study

To analyze the data, a Computer running Windows, equipped with 8GB of RAM, was employed. Using a variety of transfer learning approaches, the features from the photo files were retrieved. Two separate attribute files were taken from the images to define the person and gender. Data sets used for training and testing were divided using the method of 10-fold cross-validation. The first step in the procedure is to partition the data into n groups. One group is utilized for testing, while the other nine groups are used for training [29]. To use the entire dataset of thyroid imaging data for both testing and training, this procedure is performed ten times with the groups being switched. The Python program's classi- fication techniques were used to test attribute files made for thyroid categorization, and the best classification success was sought after [30]. The study's transfer learning techniques, classification algorithms, and success rates are discussed in the Results section. The study's process is displayed below. Fig. (1) depicts various steps followed for classifying thyroid nodules.

4.3. Classification

4.4. Transfer Learning

Transfer learning is a stream in machine learning research that deals with storing knowledge gained while solving one problem and applying it to other related problems [34]. A person who can ride a bike, for instance, could find it easier and quicker to learn how to use a scooter. When confronted with similar circumstances, the person uses their ability to stay balanced while driving a scooter and unintentionally applies what they have learned. The capacity to retain information learned from solving one problem and use it to solve another later is known as transfer learning. Models that make better use of existing knowledge and learn more quickly with less training data are developed using transfer learning. The most advantageous feature of transfer learning is that a fraction of the trained model needs learning in order to use it. We can achieve this and save time by using transfer learning.

4.5. Splitting of Data for Training and Testing

Machine learning employs the data split technique, dividing available data into training and testing sets to assess a model's performance on new data. The training set is utilized for model development, while the testing set evaluates its performance. Although a common practice involves a 70% training and 30% testing split, this ratio may vary based on dataset size and model complexity. Evaluating the model on a distinct testing set helps gauge its generalization to new, untested data. A model exhibiting good performance on the testing set is more likely to generalize well to new data, while poor performance may indicate overfitting to the training set, necessitating model adjustments. Fig. (3) shows the 10-fold validation of training and test data.

4.6. Performance Evaluation Metrics

The performance analysis of thyroid categorization systems is done using the confusion matrix. The columns of the confusion matrix, a 2x2 matrix, contain the real samples that should belong to each class, while the rows reflect the estimated sample data for each class. The row and column information can also be inverted when creating the confusion matrix. Precision, Recall, Accuracy, and F1-score rates are scores derived by the confusion matrix in Table 2 below.

Metrics for performance evaluation are used to evaluate how well a machine learning model is performing. These measures are used to assess how successfully the model generalizes to new, untrained data and learns from the training data. The challenge at hand and the model's aims influence the choice of metrics. Here are some common machine learning performance evaluation metrics:

4.6.1. Accuracy

Accuracy is determined by the percentage of correct predictions made by the model. Calculate accuracy by dividing the total number of correct predictions by the total number of assumptions (Eq 5).

(5)

4.6.2. Precision

The fraction of accurate positive forecasts overall true positive is how precision is calculated. When the objective is to reduce false positives, it is employed (Eq 6).

(6)

4.6.3. Recall

Recall measures the proportion of accurate positive predictions relative to all actual positive occurrences. It is employed when the objective is to minimize false negatives (Eq 7).

(7)

4.6.4. F1 Score

The F1 score serves as the harmonic mean of precision and recall, providing a combined metric when both precision and recall are of significance.

(8)

In both the training and testing phases, a model's performance is evaluated using these metrics. They can be used to identify areas where the model is performing poorly and to adjust the model's parameters to improve its performance.

5.1. Hyperparameter Tuning and Optimization Process

To ensure the optimal performance of the models, we performed comprehensive hyperparameter tuning and optimization. The following hyperparameters were considered for tuning (Table 4).

A grid search strategy was employed for systematic hyperparameter tuning, involving an exhaustive search over specified hyperparameter values. The primary evaluation metric used for hyperparameter selection was validation accuracy. Additionally, precision, recall, and F1 scores were monitored to ensure balanced performance across different metrics. A 5-fold cross-validation method was used to evaluate the performance of different hyperparameter configurations, ensuring that the selected hyperparameters generalize well across different data subsets. For DenseNet169, the optimal configuration included a learning rate of 0.001, batch size of 32, Adam optimizer, and a dropout rate of 0.5. For ResNet101, the best results were obtained with a learning rate of 0.0001, batch size of 64, RMSprop optimizer, and a dropout rate of 0.4. For EfficientNetB1, the optimal configuration involved a learning rate of 0.0005, batch size of 32, SGD optimizer, and a dropout rate of 0.4.

Table 5 above displays the outcomes of the DenseNet169 properties. The classification methods were evaluated based on several performance metrics, including accuracy, recall, precision, and F1 score. Discriminant Analysis Linear Discriminant Analysis (LDA) emerged as the top-performing classifier with an accuracy of 92.46%, outperforming other methods. LDA also demonstrated high recall, precision, and F1 score, indicating its robustness in correctly identifying both positive and negative instances. Linear Model Ridge Classifier and Linear Model Ridge Classifier with cross-validation (CV) achieved competitive accuracy scores of 90.92% and 90.01%, respectively, while maintaining balanced performance across other metrics. Support Vector Machine (SVM) with linear kernel and Linear Model Logistic Regression with CV also showed strong performance, with accuracy scores above 88%. Ensemble methods, such as the Ensemble Hist Gradient Boosting Classifier, while slightly lower in accuracy at 81.54%, still exhibited respectable performance across all metrics. Neural Network MLP Classifier and Linear Model Logistics Regression lagged behind in accuracy but maintained moderate performance in other metrics. Overall, Discriminant Analysis LDA stands out as the preferred classifier due to its superior accuracy and balanced performance across all evaluated metrics Fig. (5).