Deep Learning for Thyroid Nodule Classification

Abstract— Thyroid ultrasound imaging is a crucial Feature, and Homogeneity are gathered to train the
diagnostic method for assessing thyroid conditions, especially classification models. The system's performance is compared
for distinguishing benign nodules from malignant ones. This with ground truth images provided by radiologists, and
paper introduces a deep learning method focused on automating accuracy serves as the main performance measure. Results
the classification of thyroid ultrasound images, supporting
healthcare professionals in providing precise diagnoses and
show that the segmentation performance of ELM
making better treatment decisions. The method is based on outperforms the SVM classifier in this task [1].
Convolutional Neural Networks (CNNs), utilizing a ResNet The discussion about various thyroid disease by
architecture to categorize images into three distinct groups: Alyas, Tahir et al. such as hypothyroidism, hyperthyroidism,
Normal Thyroid, Malignant, and Benign. A dataset of annotated and thyroid cancer, that affect people globally. These
thyroid ultrasound images is used to train the model, with data conditions can lead to severe symptoms, making early
augmentation techniques employed to enhance its ability to detection crucial for effective treatment. Machine learning
generalize. evaluated through important metrics like accuracy, and classification techniques play a crucial role in the timely
loss, precision, recall, F1-score, and support. This approach identification of thyroid diseases, enabling more efficient
seeks to offer a reliable and efficient solution for thyroid nodule
classification, ultimately improving diagnostic precision and
patient management. Accurate identification and
contributing to better management of thyroid diseases. categorization of thyroid nodules in ultrasound images help
alleviate radiologists' workload and reduce the likelihood of
Keywords— Thyroid Ultrasound Imaging, Convolutional errors. This research evaluates various machine learning
Neural Networks (CNNs), ResNet, Thyroid Nodule approaches, including decision trees, random forests, KNN,
Classification, Diagnostic Accuracy. and neural networks. Through dataset manipulation, the
random forest algorithm demonstrated the highest accuracy
I.INTRODUCTION of 94.8% and a specificity of 91%, emphasizing its
Thyroid disorders, including both benign and malignant effectiveness in predicting thyroid conditions [2].
nodules, are prevalent conditions that can have serious health
The survey study on ultrasound imaging by Yi-
consequences if not properly diagnosed and managed. Precise
differentiation between benign and malignant nodules is crucial for Cheng Zhu et al. explain that is commonly used for
proper treatment planning, as it influences decisions related to diagnosing thyroid nodules, and AI-powered solutions are
surgery, excision, or other medical interventions. However, the gaining traction in this area. This study proposed a deep
interpretation of these images can be subjective, often relying on the learning-based algorithm to differentiate benign and
clinician's experience, which may lead to inconsistencies in malignant thyroid nodules using ultrasound images. A dataset
diagnosis. In recent years, deep learning has brought significant comprising 592 patients with 600 thyroid nodules was
advancements to medical imaging, offering powerful tools for utilized for training, validation, and testing, with an
automating image analysis and classification. Among these, additional 187 patients and 200 nodules included in an
Convolutional Neural Networks (CNNs) have shown impressive
external test set. The authors improved the Visual Geometry
performance across a variety of image classification challenges,
including applications in healthcare. These models can Group (VGG)-16 architecture by incorporating batch
automatically learn intricate patterns and features from large normalization and dropout layers, resulting in the VGG-16T
datasets, making them particularly useful for identifying normal, model. The model’s performance was assessed using a 10-
malignant, and benign thyroid nodules in ultrasound scans. fold cross-validation approach. The findings, based on
grayscale ultrasound images from five distinct machine
brands, highlighted the VGG-16T model’s robustness in
II. RELATED WORDS accurately classifying thyroid nodules [3].
The research study by D. Selvathi, V.S. Sharnitha et The analysis of thyroid nodule by Yang Jingya et al.
al., shows that the clinical reports often provide are characterized by abnormal growth in thyroid cells, which
morphometric data by comparing it to previous studies. To may indicate various conditions like excessive iodine intake,
convey clinically relevant data regarding an object’s size and degeneration, or inflammation. Though most nodules are
shape, alongside image segmentation and classification, are benign, the risk of malignancy increases yearly. To alleviate
critical aspects of medical image processing. Ultrasound, as the burden on clinicians and avoid unnecessary fine-needle
a versatile imaging method, can clearly reveal internal organ aspirations or surgeries, deep learning techniques have been
structures, often with remarkable detail. It has the unique explored for diagnosing thyroid nodules via image
capability of imaging both anatomy and physiology in real recognition. In this study, 508 ultrasound images from Hebei
time, providing a rapid, non-invasive method of assessment. Medical University were used to train a ResNet18 model. A
Ultrasound remains one of thefrequently used diagnostic 10-fold random sampling approach was applied to the
methods in clinical environments, significantly aiding patient training dataset to prevent errors. The model exhibited strong
care by providing physicians with comprehensive anatomical performance, achieving an average AUC of 0.997, accuracy
insights. In this work, an automated system is proposed for of 0.984, recall of 0.978, precision of 0.939, and F1-score of
classifying thyroid images and segmenting the thyroid gland 0.957. Additionally, the study used Gradient-weighted Class
using machine learning techniques. Classifiers like SVM and Activation Mapping (Grad-CAM) to identify key regions in
ELM are employed, and features such as mean, variance, the images during the learning process. This technique
Coefficient of Local Variation, Histogram Feature, NMSID

XXX-X-XXXX-XXXX-X/XX/$XX.00 ©20XX IEEE

revealed shape feature differences between benign and differences between layers, rather than attempting to directly
malignant nodules, proving helpful in diagnosis [4]. learn the output. This approach helps address the vanishing
According to M. Segni et al., thyroid diseases, gradient issue, allowing for more effective training of deeper
including conditions like hypothyroidism, hyperthyroidism, models. The use of residual connections not only enhances
and thyroid cancer, are prevalent worldwide, often leading to feature extraction but also ensures that performance does not
serious symptoms if left untreated. Machine learning degrade as the model becomes more complex. This
techniques and effective classification are key to the early architectural design proves particularly effective for
detection of these disorders, significantly impacting patient analyzing intricate medical images, such as thyroid
outcomes by enabling timely treatment. Automating the ultrasound scans, where subtle differences are crucial for
detection of thyroid nodules in ultrasound images can help diagnosis. ResNet-50's capacity to capture fine details makes
reduce radiologist workload and minimize diagnostic errors, it highly suitable for differentiating between various thyroid
making it an essential tool in the clinical environment [5]. conditions.

III. METHODOLOGY D. Training and Testing

The dataset for this project is carefully divided into
two key components: a training set and a testing set. The
training set, which contains a diverse collection of labeled
thyroid ultrasound images, is utilized to train the ResNet-50
model to recognize patterns and features that correspond to
different thyroid conditions. During the training process, the
model fine-tunes its internal parameters using
backpropagation and optimization methods in order to reduce
classification errors. The testing set, consisting of images that
A. Data Collection the model has not seen before, is used to assess its
performance. This phase is crucial for assessing the model's
The dataset used in this study comprises thyroid
ability to apply its learned knowledge to new, unseen data.
ultrasound images that were annotated and obtained from [6].
Various performance metrics, such as accuracy, precision,
These images are categorized into three classes: Normal
recall, and F1-score, are computed to evaluate the model's
Thyroid, Malignant, and Benign. A total of 2,289 images
effectiveness.
were included in the dataset. To ensure uniformity and
quality, all images underwent preprocessing, where they were
resized to 100x100 pixels and normalized to maintain E. Classification
consistent pixel intensity. Furthermore, data augmentation Once the ResNet-50 model has been trained, it is
techniques such as rotation, flipping, and scaling were deployed to classify new thyroid ultrasound images. During
employed to artificially expand the dataset and improve the classification, the model processes each new image through
model's generalization capabilities. The dataset was split into its network, applying learned filters and residual connections
three parts: 60% for training, 20% for validation, and 20% for to extract relevant features. The model then uses these
testing. This structure supports up to six authors but remains features to predict whether the nodule in the image is
flexible for varying author configurations. malignant, benign, or represents normal thyroid tissue. This
B. Preprocessing automated classification process aids healthcare
professionals by providing a rapid and consistent evaluation
In the preprocessing step, several actions are taken to
of ultrasound images. The model's predictions can be used to
prepare the thyroid ultrasound images for the ResNet-50
model. First, all images are resized to the same size to make support diagnostic decisions, recommend further testing, or
sure they are consistent and easy for the model to process. determine appropriate treatment strategies. The accuracy and
Next, the pixel values are normalized to reduce the effects of speed of the ResNet-50 model in classifying thyroid
different lighting conditions, helping the model focus on conditions contribute to improving patient outcomes and
important features. To make the model more reliable, enhancing diagnostic efficiency in clinical settings.
techniques like rotation, flipping, and scaling are used to
create more varied images from the existing dataset. This
helps the model learn better and improves its accuracy in IV.RESULTS AND DISCUSSION
classifying thyroid nodules as normal, malignant, or benign. The performance of the ResNet-50 model in
classifying thyroid ultrasound images was evaluated using
key metrics, including accuracy, precision, recall, and F1-
C. ResNet Model score. Following training on the preprocessed dataset, the
model was tested on a separate validation set to assess its
In this study, after undergoing preprocessing, the effectiveness. The model achieved an accuracy of 71%,
thyroid ultrasound images are processed using the ResNet-50 correctly classifying 71% of the thyroid images into the
architecture, a deep learning model known for its superior categories of malignant, benign, or normal. This result
ability to extract features. ResNet-50 incorporates residual highlights the model’s ability to distinguish between different
connections, commonly referred to as skip connections, thyroid conditions based on ultrasound images.The learning
which enable the model to focus on learning the residuals, or progress was visualized through performance graphs,
showing accuracy and loss over multiple training epochs. TABLE 1. CLASSIFICATION TABLE
Analyzing these trends provides insights into model Classification Precision Recall F1- Support
behavior, such as patterns of convergence or any signs of Score
overfitting or underfitting, which can be crucial for refining
Malignant 0.87 0.04 0.08 292
the training process and improving performance.
Precision represents the ratio of correctly identified positive Benign 0.28 0.08 0.13 60
cases among those predicted as positive, while recall Normal 0.02 0.86 0.04 7
measures the proportion of actual positive cases that were Method
correctly identified. Evaluating precision and recall in
addition to accuracy provides a more comprehensive
understanding of the model's performance, especially in V. CONCLUSION
situations with imbalanced data or when misclassification
consequences vary in severity. In this study, we examined the use of ResNet-50 for thyroid
classification, utilizing deep learning to enhance diagnostic
accuracy. By comparing the performance of the SGD and
Understanding the clinical impact of the model's results is
Adam optimizers, we found that each optimizer offers
also critical. False positives and false negatives in medical
distinct advantages. The SGD optimizer ensured more stable
diagnosis can lead to serious consequences, such as training with controlled convergence, while the Adam
unnecessary procedures or missed diagnoses, affecting optimizer facilitated faster convergence. These results
patient outcomes. Thus, evaluating these errors is essential to highlight the significant impact of optimizer selection on the
determine the model's suitability for practical clinical performance and efficiency of deep learning models in
applications.When comparing the model's performance to medical imaging. Overall, the ResNet-50 model showed
other studies, such as Yang et al., this model achieved a promise as a valuable tool for thyroid classification,
precision of 0.87, outperforming Yang et al.'s 0.84. The recall providing a reliable method to assist healthcare professionals
of 0.86 also surpasses Yang et al.'s recall of 0.77, suggesting in making informed decisions. Future research could focus on
that this model has higher sensitivity and is more effective at further fine-tuning the model, exploring alternative
accurately identifying positive thyroid nodule cases. architectures, or incorporating more diverse datasets to
improve the model's generalizability and robustness for
clinical use.
REFERENCES

[1] Prathibha, S., Dahiya, D., Robin, C. R., Nishkala, C. V.,

& Swedha, S. (2023). A novel technique for detecting
various thyroid diseases using deep learning. Intell.
Autom. Soft Comput, 35(1), 199-214.
[2] Zhao, X., Shen, X., Wan, W., Lu, Y., Hu, S., Xiao, R., ...
& Li, J. (2022). Automatic thyroid ultrasound image
classification using feature fusion network. IEEE
Access, 10, 27917-27924.
[3] Selvathi, D., and V. S. Sharnitha. "Thyroid classification
and segmentation in ultrasound images using machine
learning algorithms." 2011 International Conference on
Signal Processing, Communication, Computing and
Fig.2. Predicted Output Networking Technologies. IEEE, 2011.
[4] Vadhiraj, Vijay Vyas, et al. "Ultrasound image
classification of thyroid nodules using machine learning
techniques." Medicina 57.6 (2021): 527.
[5] Yang, Jingya, et al. "Ultrasound image classification of
thyroid nodules based on deep learning." Frontiers in
Oncology 12 (2022): 905955.
[6] Zhu, Yi-Cheng, et al. "Thyroid ultrasound image
classification using a convolutional neural network."
Annals of translational medicine 9.20 (2021).
[7] Guo, Yinxiang, Jianing Xu, Xiangzhi Li, Lin Zheng, Wei
Pan, Meiting Qiu, Shuyi Mao, Dongfei Huang, and
Xiaobo Yang. "Classification and diagnosis of residual
thyroid tissue in SPECT images based on fine-tuning
deep convolutional neural network." Frontiers in
Oncology 11 (2021): 762643.
[8] Chan, Wai-Kin, Jui-Hung Sun, Miaw-Jene Liou, Yan-
Rong Li, Wei-Yu Chou, Feng-Hsuan Liu, Szu-Tah
Fig.3. Performance Graph Chen, and Syu-Jyun Peng. "Using deep convolutional
neural networks for enhanced ultrasonographic image
diagnosis of differentiated thyroid cancer."
Biomedicines 9, no. 12 (2021): 1771.
[9] Wang, Xuanqi, Bin Cao, Dejian Wei, Jing Liu, and Hui
Cao. "Diagnosis of thyroid nodules based on lightweight
 residual network." In 2021 IEEE International [20] Sun, Chao, Yukang Zhang, Qing Chang, Tianjiao Liu,
Conference on Bioinformatics and Biomedicine Shaohang Zhang, Xi Wang, Qianqian Guo, Jinpeng Yao,
(BIBM), pp. 3875-3881. IEEE, 2021. Weidong Sun, and Lijuan Niu. "Evaluation of a deep
[10] Guo, Minghui, Kangjian Wang, Shunlan Liu, Yongzhao learning‐based computer‐aided diagnosis system for
Du, Peizhong Liu, Qichen Su, and Guorong Lv. distinguishing benign from malignant thyroid nodules in
"Recognition of thyroid ultrasound standard plane ultrasound images." Medical physics 47, no. 9 (2020).
images based on residual network." Computational [21] Daniels, Kelly, Sriharsha Gummadi, Ziyin Zhu, Shuo
Intelligence and Neuroscience 2021, no. 1 (2021): Wang, Jena Patel, Brian Swendseid, Andrej Lyshchik,
5598001. Joseph Curry, Elizabeth Cottrill, and John Eisenbrey.
[11] Buddhavarapu, Vijaya Gajanan. "An experimental study "Machine learning by ultrasonography for genetic risk
on classification of thyroid histopathology images using stratification of thyroid nodules." JAMA
transfer learning." Pattern Recognition Letters 140 Otolaryngology–Head & Neck Surgery 146, no. 1 (2020)
(2020): 1-9. [22] Chen, Junying, Haijun You, and Kai Li. "A review of
[12] Alyas, Tahir, et al. "Empirical method for thyroid disease thyroid gland segmentation and thyroid nodule
classification using a machine learning approach." segmentation methods for medical ultrasound images."
BioMed Research International 2022 (2022). J. Kim, J. Computer methods and programs in biomedicine 185
E. Gosnell and S. A. Roman, "Geographic influences in (2020).
the global rise of thyroid cancer", Nature Rev. [23] Grani, Giorgio, Marialuisa Sponziello, Valeria Pecce,
Endocrinol., vol. 16, no. 1, pp. 17-29, Jan. 2020. Valeria Ramundo, and Cosimo Durante. "Contemporary
[13] H. Ye, J. Hang, X. Chen, D. Xu, J. Chen, X. Ye, et al., thyroid nodule evaluation and management." The
"An intelligent platform for ultrasound diagnosis of Journal of Clinical Endocrinology & Metabolism 105,
thyroid nodules", Sci. Rep., vol. 10, no. 1, Aug. 2020. no. 9 (2020).
[14] Wang, Yunjun, Qing Guan, Iweng Lao, Li Wang, Yi Wu, [24] Shi, Guohua, Jiawen Wang, Yan Qiang, Xiaotang Yang,
Duanshu Li, Qinghai Ji et al. "Using deep convolutional Juanjuan Zhao, Rui Hao, Wenkai Yang, Qianqian Du,
neural networks for multi-classification of thyroid tumor and Ntikurako Guy-Fernand Kazihise. "Knowledge-
by histopathology: a large-scale pilot study." Annals of guided synthetic medical image adversarial
translational medicine 7, no. 18 (2019). augmentation for ultrasonography thyroid nodule
classification." Computer Methods and Programs in
[15] Yu, Xia, Hongjie Wang, and Liyong Ma. "Detection of Biomedicine 196 (2020).
thyroid nodules with ultrasound images based on deep
learning." Current Medical Imaging 16, no. 2 (2020): [25] Xu, Lei, Junling Gao, Quan Wang, Jichao Yin, Pengfei
174-180. Yu, Bin Bai, Ruixia Pei et al. "Computer-aided diagnosis
systems in diagnosing malignant thyroid nodules on
[16] Thomas, Johnson, Gregory A. Ledger, and Chaitanya K. ultrasonography: a systematic review and meta-
Mamillapalli. "Use of artificial intelligence and machine analysis." European thyroid journal 9, no. 4 (2020).
learning for estimating malignancy risk of thyroid
nodules." Current Opinion in Endocrinology, Diabetes [26] Guo, Xiangdong, Haifeng Zhao, and Zhenyu Tang. "An
and Obesity 27, no. 5 (2020) improved deep learning approach for thyroid nodule
diagnosis." In 2020 IEEE 17th International Symposium
[17] Nugroho, Hanung Adi, Eka Legya Frannita, and on Biomedical Imaging (ISBI), pp. 296-299. IEEE,
Augustine Herini Tita Hutami. "Thyroid nodules 2020.
categorization based on margin features using deep
learning." In 2020 3rd International Seminar on Research [27] Zhang, Shijie, Huarui Du, Zhuang Jin, Yaqiong Zhu,
of Information Technology and Intelligent Systems Ying Zhang, Fang Xie, Mingbo Zhang, Xiaoqi Tian, Jue
(ISRITI), pp. 499-504. IEEE, 2020. Zhang, and Yukun Luo. "A novel interpretable
computer-aided diagnosis system of thyroid nodules on
[18] Yu, Xia, Hongjie Wang, and Liyong Ma. "Detection of ultrasound based on clinical experience." Ieee Access 8
thyroid nodules with ultrasound images based on deep (2020): 53223-53231.
learning." Current Medical Imaging 16, no. 2 (2020):
[19] Song, Ruoning, Long Zhang, Chuang Zhu, Jun Liu, Jie
Yang, and Tong Zhang. "Thyroid nodule ultrasound
image classification through hybrid feature cropping
network." IEEE Access 8 (2020).