ThyroNet-X4 Genesis: An Advanced Deep Learning Model for Auxiliary Diagnosis of Thyroid Nodules' Malignancy

ThyroNet-X4 Genesis: An Advanced Deep Learning Model for Auxiliary Diagnosis of Thyroid Nodules' Malignancy | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Article ThyroNet-X4 Genesis: An Advanced Deep Learning Model for Auxiliary Diagnosis of Thyroid Nodules' Malignancy Xiaoxue Wang, Yupeng Niu, Hongli Liu, Fa Tian, Qiang Zhang, Yimeng Wang, and 2 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-4408975/v1 This work is licensed under a CC BY 4.0 License Status: Published Journal Publication published 04 Feb, 2025 Read the published version in Scientific Reports → Version 1 posted 10 You are reading this latest preprint version Abstract Thyroid nodules are common endocrine disorders, and the discrimination between benign and malignant nodules is crucial for treatment decisions. Traditional ultrasound diagnosis relies on the experience of physicians, which may pose risks of misdiagnosis. In this study, we propose a novel deep learning model, ThyroNet-X4 Genesis, for the automatic classification of thyroid nodules' malignancy. This model is based on the ResNet module, which optimizes computational efficiency and enhances feature extraction capabilities by introducing grouped convolution and increasing the convolution kernel size, thus extracting features and classifying nodules in ultrasound images. We obtained data from publicly available medical imaging databases for internal training and validation and used ultrasound images collected from HanZhong Central Hospital as an external validation set to evaluate the model's generalization ability and practical application value.ThyroNet-X4 Genesis achieved training and validation accuracies of 85.55% and 71.70%, respectively, on the internal validation set, with a testing accuracy of 67.02% on the external validation set, outperforming other mainstream comparative models, indicating its good performance in actual clinical applications. The development of this model showcases the potential of deep learning in thyroid imaging analysis, providing valuable references for future development of high-performance medical diagnostic models. Health sciences/Endocrinology Health sciences/Endocrinology/Endocrine system and metabolic diseases Health sciences/Endocrinology/Endocrine system and metabolic diseases/Thyroid diseases Thyroid nodules Deep learning Auxiliary diagnosis External validation Clinical interpretability Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 1. Introduction Thyroid nodules are defined as space-occupying lesions within the thyroid gland that can be detected through imaging studies, differentiated from the surrounding thyroid tissue. These nodules can be benign or malignant, and according to recent epidemiological studies, up to 60% of the population have thyroid nodules[ 1 ], with a malignancy rate of approximately 1%-5%[ 2 ]. Further research indicates that the prevalence of thyroid nodules over 0.5 cm in diameter is 20.43%, with 8%-16% being malignant[ 3 ]. The treatment approaches for thyroid nodules vary based on their nature; benign nodules often require no treatment but regular follow-up. However, surgical intervention becomes necessary when nodules grow large enough to cause compressive symptoms such as difficulty breathing, swallowing difficulties, and hoarseness. Malignant thyroid nodules, posing a threat to the patient's life and quality of life, require accurate diagnosis and surgical treatment, making the differentiation of their nature a core aspect of assessment. In the general population, the incidence of palpably detected thyroid nodules is between 3% and 7%, but with the aid of high-resolution ultrasound, detection rates can soar to between 20% and 76%[ 4 ]. Compared to other diagnostic modalities like X-rays, MRI, and CT, ultrasound offers advantages such as efficiency, convenience, and the absence of radiation. With advancements in ultrasound resolution, technologies such as ultrasound contrast enhancement and elastography have rapidly evolved, making color Doppler ultrasound the preferred method for diagnosing thyroid nodules[ 5 ]. In 2011, Russ and colleagues[ 6 ] used indicators such as very low echogenicity, microcalcifications, an aspect ratio > 1, and irregular margins or borders to develop a five-tier thyroid imaging reporting and data system (TIRADS), assessing the malignancy risk of thyroid nodules and facilitating the identification and further management of potentially malignant nodules. Subsequently, South Korea, Europe, the United States, and China have progressively established their own TIRADS[ 7 – 10 ], which similarly use solid composition, low echogenicity, irregular margins, vertical growth, and microcalcifications as indicators for suspicious malignancy in assessing and grading the risk of thyroid nodules. However, the process of describing ultrasound characteristics of thyroid nodules and quantifying risk levels according to TIRADS standards can be time-consuming and may vary in accuracy due to the experience of the ultrasound physicians and the quality of the diagnostic equipment. With the rise of artificial intelligence, more studies are employing deep learning for ultrasound detection of thyroid nodules. Chi[ 11 ] and others proposed the GoogLeNet model, extracting features from thyroid ultrasound images and inputting these into a random forest classifier to distinguish between benign and malignant thyroid nodules. Wang[ 12 ] and others improved the Faster RCNN model to better extract ultrasound features of thyroid papillary carcinoma, enhancing diagnostic accuracy. Liang[ 13 ] developed a deep learning model specifically for classifying thyroid and breast nodules. Zhang[ 14 ] utilized the YOLOv3 model to discriminate between benign and malignant thyroid nodules in TIRADS category 4, significantly impacting subsequent treatment decisions and patient outcomes. Moussa[ 15 ] used the ImageNet-pretrained ResNet50 for transfer learning, achieving promising diagnostic results in their own ultrasound image dataset. Kwon[ 16 ] employed a pretrained VGG16 model for transfer learning, effectively classifying thyroid nodules based on malignancy. Clearly, deep learning holds significant value in enhancing the accuracy of thyroid nodule diagnoses, reducing physician workload, and standardizing diagnostic procedures. However, most current studies focus on single-modality ultrasound images or are only internally validated on a single dataset, with further improvements needed in accuracy, model generalization, and stability. In this study, we innovatively propose the CNN-based ThyroNet-X4 Genesis model, initially trained and cross-validated using publicly available database data, while also collecting ultrasound imaging and related clinical data from our center as an external validation set to assess the model's generalization ability and practical value. By increasing the expansion factor and the size of the convolutional kernels without adding computational burden, we significantly enhanced the model's expressive capacity and its ability to extract ultrasound features. Additionally, we incorporated grouped convolutions as an effective method to reduce the number of parameters and computational complexity, enhancing the model's learning ability across different feature channel groups. The ThyroNet-X4 Genesis model achieved optimal balance across all configurations, exhibiting the lowest training losses and the highest training and validation accuracies, effectively boosting the model's generalization capabilities and demonstrating the importance of innovative network design ideas in enhancing the performance of deep learning models. The workflow of this work is shown in Fig. 1 . 2. Results 2.1 Internal Validation Results 2.1.1 ThyroNet-X4 Genesis Model Results In our study, the ThyroNet-X4 Genesis model demonstrated excellent performance in internal validation, indicating its efficiency and accuracy in the task of differentiating benign and malignant thyroid nodules. We utilized the Adam optimizer with an initial learning rate set to 0.001 and implemented a learning rate decay strategy to finely tune the training process. The batch size was adjusted to 32 based on experimental settings and hardware configuration, ensuring sufficient data loading without excessive resource consumption. During the validation phase, the ThyroNet-X4 Genesis model exhibited superior generalization capability. Figure 2 illustrates the relevant results over 100 training epochs. Table 1 presents the specific training and validation results of the model. In this table, the ThyroNet-X4 Genesis model achieved excellent training loss and training accuracy across all configurations. Additionally, it obtained relatively high accuracy on the validation set, indicating that the model not only learns effectively but also possesses good predictive ability on unseen data. Table 1 Results of ThyroNet-X4 Genesis model Model train_loss train_accuracy val_loss val_accuracy ThyroNet-X4 Genesis(best) 0.2595 0.8555 0.6149 0.7170 2.1.2 Comparative Model Results In this study, through a series of ablation experiments, we extensively explored the performance differences among various model variants to assess the impact of structural adjustments (as shown in Table 2 and Fig. 3 ). The baseline model exhibited a relatively high training accuracy (84.926%) but a relatively lower validation accuracy (69.5418%), implying possible overfitting. The improved Block variant (Baseline + Block) demonstrated better generalization capability with slightly lower training accuracy (84.7028%) and an increased validation accuracy of 70.6199%, suggesting that the additional Block structure helps improve the model's performance on unseen data. Although the Bottleneck variant (Baseline + Bottleneck) achieved the highest training accuracy (85.237%), the validation accuracy decreased to 68.4636%, indicating exacerbated overfitting. Table 2 Performance of each comparison model Model train_loss train_accuracy val_loss val_accuracy Baseline 0.2795 0.8493 1.393 0.6954 Baseline + Bottleneck 0.2740 0.8524 1.1411 0.6846 Baseline + Block 0.2751 0.8470 0.6972 0.7061 VGGNet 0.6094 0.6694 1.1586 0.5983 AlexNet 0.5822 0.6561 2.2945 0.4952 GoogleNet 0.9523 0.6728 1.6618 0.5705 MobileNet 0.5803 0.6885 1.6088 0.5971 The ThyroNet-X4 Genesis model showed the optimal balance across all configurations, with the lowest training loss (0.259516), the highest training (85.5478%), and validation accuracy (71.6981%), significantly enhancing its generalization capability. Additionally, comparisons were made with other deep learning architectures such as VGGNet, AlexNet, GoogleNet, and MobileNet. VGGNet performed well in extracting shape and texture features due to its deep but simple convolutional structure, yet it exhibited the lowest validation accuracy (59.83%), suggesting its potential inadequacy in handling the complexity of thyroid nodules. AlexNet and MobileNet showed relatively lower validation accuracy (49.52% and 59.71%, respectively). Although they are efficient in resource-constrained environments, they have limitations in handling fine-grained features. GoogleNet, with its Inception module effectively capturing features at different scales, still had a low validation accuracy (57.05%), indicating that its complex structure might be overly intricate for the dataset used in this study. 2.2 External Center Validation and Clinical Interpretability To further evaluate the potential clinical application of the ThyroNet-X4 Genesis model, we applied it to ultrasound images of 658 thyroid nodule patients from our center (with a test accuracy of 67.02%) and conducted a detailed analysis of its clinical interpretability. We employed Grad-CAM (Gradient-weighted Class Activation Mapping) technology to generate heatmaps, highlighting the regions of interest the model focused on (as shown in Fig. 4 ). This visualization technique helps doctors verify the model's rationale and ensure that its focus aligns with the clinical diagnostic process. To further assess the model's clinical interpretability, we invited three ultrasound doctors with years of thyroid diagnosis experience to independently evaluate the model's diagnostic results. The doctors assessed the model's diagnostic results based on their professional knowledge and pointed out whether the features the model focused on in certain cases aligned with clinical diagnostic criteria. The results showed that in most cases, the regions of interest identified by ThyroNet-X4 Genesis were consistent with those identified by the doctors, and the model's diagnostic accuracy for malignant nodules was comparable to that of the doctors. Furthermore, we conducted in-depth analysis of misdiagnosed and missed cases by the model to assess its potential clinical risks. We found that in certain types of thyroid nodules, such as mixed nodules and ectopic thyroid nodules, the model's diagnostic accuracy was relatively low. These types of nodules often exhibit complex morphological features in ultrasound images, making it easy for the model to confuse benign and malignant conditions. In such cases, the model's diagnostic results can serve as supplementary information for doctors rather than the sole basis for diagnosis. 3. Discussion In this study, we proposed a novel deep learning model, ThyroNet-X4 Genesis, which trained on thyroid ultrasound images from publicly available medical imaging databases to classify the benign and malignant nature of thyroid nodules. We used ultrasound images from our center as the validation set, and the results showed that the ThyroNet-X4 Genesis model exhibited the lowest training loss (0.259516), highest training accuracy (85.5478%), and highest validation accuracy (71.6981%) across all configurations, demonstrating its superior generalization capability. Several reasons contribute to the optimal balance demonstrated by this model. Firstly, the expansion coefficient was increased from 1 to 4, significantly enlarging the channel number of the network's output layer. This improvement allowed the network to carry more information without significantly increasing computational burden, thus significantly enhancing the model's expressive power. Secondly, improvements were made in convolution kernel size and grouped convolution. We used a 5x5 convolution kernel in the model's second convolutional layer to capture broader contextual information and enhance feature extraction capabilities. Additionally, introducing grouped convolution as an effective method to reduce parameter count and computational complexity strengthened the model's ability to learn from different feature channel groups. Lastly, the model's training set sourced from publicly available medical imaging databases, indicating that the model learned from ultrasound images from different medical centers, thus enhancing its generalization. The core of thyroid nodule assessment lies in distinguishing between benign and malignant nodules. Thyroid ultrasound is the preferred examination for assessing nodule malignancy risk based on ultrasound image features and graded according to the TIRADS criteria. However, this process relies on the experience of ultrasound doctors and ultrasound equipment conditions, leading to a certain degree of misdiagnosis. In recent years, with the continuous advancement of deep learning technology, an increasing number of models have been developed for the diagnosis of thyroid nodule malignancy, improving the efficiency and accuracy of thyroid nodule diagnosis. For instance, Guan et al. [ 20 ] used the InceptionV3 model to classify thyroid ultrasound images, achieving high sensitivity and specificity in the test group. Ma et al. [ 21 ] fused two CNN-based models for classifying thyroid nodule malignancy, achieving an internal accuracy of up to 83.02%. PENG et al. [ 22 ] developed the ThyNet model, which showed significantly higher diagnostic accuracy than professional doctors. However, most studies to date have not applied the models to external data for validation, limiting the models' universality and clinical application value. In this study, the ThyroNet-X4 Genesis model achieved a training accuracy of up to 85.6%, validation accuracy of 71.7%, and external validation accuracy of 67.0%, indicating its potential for widespread application. However, this study also has some limitations. Firstly, the ThyroNet-X4 Genesis model proposed in this study only identifies the benign and malignant types of thyroid nodules, without classifying benign lesions such as inflammation, cysts, and adenomatous nodules. The included malignant thyroid nodule pathology results were all thyroid papillary carcinomas, and rare types of thyroid cancer such as medullary thyroid carcinoma and thyroid follicular carcinoma were not included. Future research can further classify these types. Additionally, the model was only validated at our center, and the provided malignant thyroid nodule cases were relatively few, thus the validation accuracy may be underestimated. Lastly, including data from more centers to increase the size of the training set could further improve the diagnostic accuracy of the model. In conclusion, our proposed ThyroNet-X4 Genesis model, which incorporates an increase in expansion coefficient and optimized convolutional layer configuration, has successfully improved the performance of the model in the task of benign and malignant diagnosis of thyroid nodules. These improvements not only enhanced the model's expressive power but also increased its generalization capability, providing valuable insights for the application of deep learning models in medical image analysis. Through comprehensive analysis, our best model demonstrated significant advantages in balancing computational efficiency and diagnostic accuracy, proving the importance of innovative network design concepts in enhancing the performance of deep learning models. 4. Methods 4.1 Data Acquisition 4.1.1 Inclusion and Exclusion Criteria The inclusion and exclusion criteria for this study are as follows. Inclusion criteria include: 1. Thyroid ultrasound assessments conducted prior to surgery or biopsy, 2. Definitive histopathological results obtained after surgery or biopsy, 3. Ultrasound images of thyroid nodules that include complete transverse and longitudinal sectional views. Exclusion criteria include: 1. Indeterminate pathological results, 2. Ultrasound images that do not display the entire extent of the nodules, patients with a history of invasive treatments such as surgery or ablation, 3. Ultrasound images that are unclear or obscured by ultrasound marker lines, blood flow signals, etc. 4.1.2 Data Collection In this study, data collection was conducted in two parts. Firstly, we utilized the publicly accessible medical imaging database DDTI (Digital Database of Thyroid Ultrasound Images), supported by the National University of Colombia, CIM@LAB, and IDIME (Institute of Medical Diagnostics). This database currently includes 299 cases and has been expanded to contain 620 benign and 914 malignant thyroid nodules, totaling over 1500 ultrasound images. Each case is presented as an XML file containing expert annotations and patient information. The database is regularly updated with new cases and images for the development of computer-aided diagnostic systems and serves as a training and teaching tool for new radiologists. These data are used for the initial training and cross-validation of the model. Secondly, we collected thyroid ultrasound images and related clinical data from Hanzhong Central Hospital in Shaanxi Province, conducting a retrospective analysis of data from 155 patients treated at our center, including 55 benign nodules and 100 malignant nodules, totaling 658 ultrasound images (Fig. 5 ). This dataset did not participate in the initial model training but served as an external validation set to assess the model's generalization ability and practical application value. It is noteworthy that this retrospective study was conducted in accordance with the Declaration of Helsinki and was approved by the Ethics Committee of Hanzhong Central Hospital (Approval No.2024(18)), and waived the requirement for the written informed consent of the patients, because the selected clinical and imaging data in this retrospective study would not affect the prognosis and privacy of the patients. 4.2 Data Preprocessing The collected images underwent preprocessing, which included steps such as denoising, normalization, and resizing to a uniform size to ensure the quality and consistency of the image data input into the model. Additionally, techniques such as rotation, scaling, and mirroring were employed to enhance the diversity of the image data and improve the model's generalization ability. 4.3 Construction of Diagnostic Model 4.3.1 Proposal of a Deep Learning-Based Model Deep learning models facilitate the extraction and classification of benign and malignant characteristics from ultrasound images of thyroid nodules. U-NET-based models are widely used in the diagnosis of thyroid nodules; Wu[ 17 ] and others built on the U-NET to introduce a method for ultrasound image segmentation of thyroid nodules based on joint upsampling, achieving precise localization of the target thyroid. However, this model is more complex than the U-NET, resulting in longer computation times. The ReAgU-Net model proposed by Ding[ 18 ] increases the backward propagation gradient to address the loss of spatial information due to increased network depth, although its performance diminishes when the contrast between nodules and background is low. To enhance the accuracy of AI diagnostics of thyroid nodules, Wei[ 19 ] and others improved upon DenseNet to develop a precise post-localization integrated deep learning classification model for thyroid nodules, though this model does not analyze a wide range of thyroid nodule pathologies and only provides classification results without standards or texture analysis. In this study, we innovatively propose the ResNet-based ThyroNet-X4 Genesis model. The ThyroNet-X4 Genesis model developed in this research is a deep convolutional neural network (CNN) specifically designed for the automatic identification and classification of the benign and malignant nature of thyroid nodules. The model integrates multiple layers of convolutional and pooling layers, utilizing 3x3 and 5x5 convolutional kernels to intricately capture the shape, edges, and texture of the nodules. To enhance the learning efficiency of the deep network and prevent issues of gradient vanishing, the model incorporates techniques from residual networks (ResNet) and densely connected networks (DenseNet), which bolster feature transmission and reuse through residual and dense connections, respectively. Moreover, an expansion coefficient was introduced before the output layer, increasing from 1 to 4, which widens the network, enabling it to handle more information without significantly increasing the computational burden. The use of grouped convolution techniques also helps to reduce the number of parameters and computational complexity. These innovative designs have resulted in exceptional performance of the ThyroNet-X4 Genesis during initial training and cross-validation on public medical imaging databases, as well as demonstrating outstanding generalization capability and high accuracy on the external validation set at our center. Specific architectural and parameter details of the model can be found in Fig. 6 (Model Architecture Diagram). The ThyroNet-X4 Genesis model is equipped with a multi-layer convolutional network structure that integrates residual and dense connectivity technologies, as well as optimized computational efficiency through expansion coefficients and grouped convolution. These features enable it to excel in the diagnosis of benign and malignant thyroid nodules. Similarly, these characteristics are applicable to the task of identifying breast tumors, as the diagnosis of breast tumors requires in-depth analysis of complex image features such as irregular margins and echo heterogeneity, which are also common in thyroid nodule images. The advanced feature processing capabilities and excellent generalization ability of the ThyroNet-X4 Genesis model demonstrate its effective adaptability for the recognition and classification of breast tumors, showcasing its potential for broad application in the field of medical image analysis. 4.3.2 Comparative Models In our study, the performance of the ThyroNet-X4 Genesis model was thoroughly evaluated by comparing it against a range of deep learning architectures. These included the traditional ResNet34, which served as a baseline model and has demonstrated excellent performance in medical image processing due to its residual network architecture. Additionally, the enhanced Baseline + Block variant aimed to improve diagnostic accuracy by strengthening local feature processing, while the Baseline + Bottle variant introduced bottleneck layers to optimize the use of computational resources and speed up processing. VGGNet is known for its deep, simple convolutional structure that excels in extracting image textures and shapes. Although simpler in structure, AlexNet is notable for its effectiveness in rapid preliminary feature extraction. GoogleNet, with its Inception modules, captures image features effectively at various scales, making it well-suited for handling complex thyroid nodule image data. MobileNet demonstrates efficient performance under resource-constrained conditions, especially suitable for quickly processing large volumes of data. These comparisons not only highlighted the superior performance of ThyroNet-X4 Genesis in diagnosing thyroid nodules but also emphasized its efficiency and potential in handling complex medical image data, proving its effectiveness and innovativeness as a diagnostic tool for thyroid nodules. 4.4 Experimental Setup In this study, we acquired thyroid ultrasound image data from a public database and precisely segmented it into training and validation sets. Specifically, after image processing and data augmentation, the training set included 622 malignant and 624 benign images, while the validation set comprised 292 malignant and 286 benign images. Additionally, data collected from Hanzhong Central Hospital served as an external test set to further validate the model's generalization ability and practical application effectiveness, containing 192 benign and 410 malignant thyroid ultrasound images. The experiments were conducted on an Ubuntu 20.04 operating system, programmed using Python 3.8, and primarily utilizing PyTorch 1.10.0 as the deep learning framework, with computational acceleration provided by CUDA 11.3. In terms of hardware, our laboratory was equipped with RTX A5000 GPUs and a server powered by an Intel(R) Xeon(R) Platinum 8358P CPU @ 2.60GHz with 15 vCPUs, ensuring ample computational resources and efficient processing capabilities. During the model training phase, we rigorously optimized all model parameters, including adjustments to learning rates and batch sizes, to ensure optimal performance on the training and validation sets and to prevent overfitting. Our tuning methods ensured the stability and reliability of the models, while independent external test sets were used to evaluate performance on unseen data, verifying their accuracy and applicability in real-world applications. 4.5 Model Evalution When evaluating deep learning models, we often use several key metrics to measure performance, including accuracy (ACC), F1 score, and so on. These evaluation metrics collectively describe the model's performance in various aspects, including prediction accuracy, comprehensiveness, and consistency between predicted and actual results. First, accuracy (ACC) is the most intuitive evaluation metric, representing the ratio of correctly classified sample data to the total number of samples. Its mathematical formula is expressed as follows: $$ACC=\frac{{(TP+TN)}}{{(TP+TN+FP+FN)}}$$ 1 Where TP represents the number of true positive samples, TN represents the number of true negative samples, FP represents the number of false positive samples, and FN represents the number of false negative samples. A higher accuracy indicates a more effective classifier and higher precision in the predicted results. Secondly, the F1 score is the harmonic mean of precision and recall. Precision represents the number of samples determined as positive examples, while recall represents the proportion of correctly predicted positive samples out of all actual positive samples. The formula for calculating the F1 score is: $$PRE=\frac{{TP}}{{(TP+FP)}}$$ 2 $$REC=\frac{{TP}}{{(TP+FN)}}$$ 3 $${F_1}=\frac{{2P * R}}{{P+R}}$$ 4 Declarations Funding This article has no funding. Acknowledgements We thank National University of Colombia, CIM@LAB, and IDIME for the open data resources and all investigators who participated in the study. Author contributions statement X.X.W and Y.P.N were responsible for the conceptualization, methodology, software, investigation, formal analysis, and writing - original draft.their contributions to this study are the same, so X.X.W and Y.P.N are the co-first authors. L.L was responsible for conceptualization, methodology, data collection, writing-original draft. T was responsible for methodology, software, investigation. Z, Y.M.W and Y.J.W were responsible for data collection. J.L was responsible for the design and review of the study, she is the co-corresponding authors of this study. All authors reviewed the manuscript. Data availability statement The datasets, from Hanzhong Central Hospital in Shaanxi Province, China, analyzed and generated in this study are not publicly available because the institution prohibits the public upload of any patient's private data. However, partial datasets are available from the corresponding author upon reasonable request, please contact [email protected] by email. Additional Information There are no conflict of interest in this study. All authors have read and approved this version of the article ,and due care has been taken to ensure the integrity of the work . Neither the entire paper nor any part of its content has been published or has been accepted elsewhere . It is not being submitted to any other journal. Ethics approval: Ethical Committee: The Biomedical Ethics Review Committee of HanZhong Central Hospital. Ethical Approval Number:No.2024（18）. This was a retrospective study, and the study data were partly derived from the publicly accessible medical imaging database DDTI (Digital Database of Thyroid Ultrasound Images), supported by the National University of Colombia, CIM@LAB, and IDIME (Institute of Medical Diagnostics) and partly from Hanzhong Central Hospital in Shaanxi Province, China. The selected clinical and imaging data was waived the requirement for the written informed consent of the patients, because the selected clinical and imaging data in this retrospective study would not affect the prognosis and privacy of the patients. Author Statement: Author 1 : Conceptualization, Methodology, Software, Investigation, Formal Analysis, Writing - Original Draft. Author 2 : Conceptualization, Methodology, Data collection, Writing-original draft. Author 3 : Methodology, Software, Investigation. Author 4,5 and 6 : Data collection. Corresponding Author: Designing and review of the study. All authors reviewed the manuscript. References Guth S Theune U Aberle J Galach A & Bamberger CM. Very high prevalence of thyroid nodules detected by high frequency (13 MHz) ultrasound examination. European Journal of Clinical Investigation 200939699–706. (10.1111/j.1365-2362.2009.02162.x) Grussendorf M Ruschenburg I & Brabant G. Malignancy rates in thyroid nodules: a long-term cohort study of 17,592 patients. European Thyroid Journal 202211 e220027. (10.1530/ETJ-22-0027) Li Y, Teng D, Ba J, et al. Efficacy and Safety of Long-Term Universal Salt Iodization on Thyroid Disorders: Epidemiological Evidence from 31 Provinces of Mainland China[ J]. Thyroid, 2020,30(4):568–579. DOI: 10. 1089 / thy. 2019. 0067. Durante C, Grani G, Lamartina L, et al. The Diagnosis and management of thyroid nodules: a review[ J]. JAMA, 2018,319 (9):914–924. DOI: 10. 1001 / jama. 2018. 0898. Gharib H, Papini E, Garber JR, et al. American Association of Clinical Endocrinologists, American College of Endocrinology, and Associazione Medici Endocrinologi Medical Guidelines for Clinical Practice for the Diagnosis and Management of Thyroid Nodules—2016 Update [ J]. Endocr Pract, 2016,22 (5):622–639. DOI: 10.4158/EP161208 . GL. Russ G,Bigorgne C,Royer B,et al.The Thyroid Imaging Reporting and Data System (TIRADS) for ultrasound of the thyroid[J].J Radiol,2011,92(7/8):701–713. Lee JY, Baek JH, Ha EJ, et al. 2020 Imaging Guidelines for Thyroid Nodules and Differentiated Thyroid Cancer: Korean Society of Thyroid Radiology[J]. Korean J Radiol, 2021,22(5):840–860.DOI: 10. 3348 / kjr. 2020. 0578. Zhou J, Yin L, Wei X, Zhang S, Song Y, Luo B, et al.. 2020 Chinese guidelines for ultrasound Malignancy risk stratification of thyroid nodules: the C-TIRADS. Endocrine. (2020) 70:256–79. doi: 10.1007/s12020-020-02441-y . Leenhardt L, Erdogan MF, Hegedus L, et al. 2013 European thyroid association guidelines for cervical ultrasound scan and ultrasound-guided techniques in the postoperative management of patients with thyroid cancer [ J]. Eur Thyroid J, 2013, 2 (3): 147–159.DOI: 10. 1159 / 000354537. Kwak JY, Han KH, Yoon JH, et al. Thyroid imaging reporting and data system for US features of nodules: a step in establishing better stratific-ation of cancer risk[ J]. Radiology, 2011,260 (3):892–899. DOI: 10. 1148 / radiol. 11110206. Chi J, Walia E, Babyn P, et al. Thyroid Nodule Classification in Ultrasound Images by FineTuning Deep Convolutional Neural Network[J]. Journal of Digital Imaging, 2017, 30(4): 477–486. Wang Y H, Ke W, Wan P. A Method of Ultrasonic Image Recognition for Thyroid Papillary Carcinoma Based on Deep Convolution Neural Network[J]. NeuroQuantology, 2018, 16(5). Liang X, Yu J, Liao J, et al. Convolutional Neural Network for Breast and Thyroid Nodules Diagnosis in Ultrasound Imaging[J]. BioMed Research International, 2020, 2020:1–9. Zhang X, Jia C, Sun M, Ma Z. The application value of deep learning-based nomograms in benign-malignant discrimination of TI-RADS category 4 thyroid nodules. Sci Rep. 2024;14(1):7878. doi: 10.1038/s41598-024-58668-6 . PMID: 38570589; PMCID: PMC10991510. Moussa O, Khachnaoui H, Guetari R, et al. Thyroid nodules classification and diagnosis in ultrasound images using fine-tuning deep convolutional neural network[J]. International Journal of Imaging Systems and Technology, 2019, 30(1):185–195. Kwon S W, Choi I J, Kang J Y, et al. Ultrasonographic Thyroid Nodule Classification Using a Deep Convolutional Neural Network with Surgical Pathology[J]. Journal of digital imaging,2020.Kumar V, Webb J, Gregory A, et al. Automated segmentation of thyroid nodule, gland, and cystic components from ultrasound images using deep learning. IEEE Access, 2020, 8: 63482–63496. Wu J, Zhang Z, Zhao J, et al. Ultrasound image segmentation of thyroid nodules based on joint up-sampling//The 2020 Second International Conference on Artificial Intelligence Technologies and Application (ICAITA), Dalian: AEIC-Academic Exchange Information Center, 2020, 1651(1): 012157. Ding J, Huang Z, Shi M, et al. Automatic thyroid ultrasound image segmentation based on U-shaped network//2019 12th International Congress on Image and Signal Processing, Biomedical Engineering and Informatics (CISP-BMEI), Suzhou: IEEE, 2019: 1–5. Wei X, Gao M, Yu R, et al. Ensemble deep learning model for multicenter classification of thyroid nodules on ultrasound images.Medical Science Monito, 2020, 26: e926096. Guan Q, Wang Y, Du J, et al. Deep learning based classification of ultrasound images for thyroid nodules: a large scale of pilot study.Annals of Translational Medicine, 2019, 7(7): 137. Ma J, Wu F, Zhu J, et al. A pre-trained convolutional neural network based method for thyroid nodule diagnosis. Ultrasonics. 2017;73:221–30. PENGS,LIUY,LV,etal.Deeplearning-basedartificialintelligencemodeltoassistthyroidnodulediagnosisand management:a multicentrediagnosticstudy[J].LancetDigit Health,2021,3(4):e250. Additional Declarations No competing interests reported. Cite Share Download PDF Status: Published Journal Publication published 04 Feb, 2025 Read the published version in Scientific Reports → Version 1 posted Editorial decision: Revision requested 02 Sep, 2024 Reviews received at journal 23 Aug, 2024 Reviews received at journal 11 Aug, 2024 Reviewers agreed at journal 05 Aug, 2024 Reviewers agreed at journal 31 Jul, 2024 Reviewers invited by journal 19 Jun, 2024 Editor assigned by journal 19 Jun, 2024 Editor invited by journal 18 May, 2024 Submission checks completed at journal 15 May, 2024 First submitted to journal 12 May, 2024 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-4408975","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":304871576,"identity":"b420e2e0-f0be-4005-b0e1-984c8d4a74ae","order_by":0,"name":"Xiaoxue Wang","email":"","orcid":"","institution":"HanZhong Central Hospital","correspondingAuthor":false,"prefix":"","firstName":"Xiaoxue","middleName":"","lastName":"Wang","suffix":""},{"id":304871577,"identity":"40c885f1-7e0d-4e7f-9ee3-de74b823ce56","order_by":1,"name":"Yupeng Niu","email":"","orcid":"","institution":"Sichuan Agricultural University","correspondingAuthor":false,"prefix":"","firstName":"Yupeng","middleName":"","lastName":"Niu","suffix":""},{"id":304871578,"identity":"85298019-5b53-496e-8bce-05bb843d19f7","order_by":2,"name":"Hongli Liu","email":"","orcid":"","institution":"HanZhong Central Hospital","correspondingAuthor":false,"prefix":"","firstName":"Hongli","middleName":"","lastName":"Liu","suffix":""},{"id":304871579,"identity":"80465e15-8c58-44a7-9908-53707f139316","order_by":3,"name":"Fa Tian","email":"","orcid":"","institution":"Sichuan Agricultural University","correspondingAuthor":false,"prefix":"","firstName":"Fa","middleName":"","lastName":"Tian","suffix":""},{"id":304871580,"identity":"36dc51ca-eae9-4b9b-bae5-e0c8282c4d92","order_by":4,"name":"Qiang Zhang","email":"","orcid":"","institution":"HanZhong Central Hospital","correspondingAuthor":false,"prefix":"","firstName":"Qiang","middleName":"","lastName":"Zhang","suffix":""},{"id":304871581,"identity":"955d0944-df18-4066-9349-1614b1b29d8e","order_by":5,"name":"Yimeng Wang","email":"","orcid":"","institution":"HanZhong Central Hospital","correspondingAuthor":false,"prefix":"","firstName":"Yimeng","middleName":"","lastName":"Wang","suffix":""},{"id":304871582,"identity":"20d5ea74-ff4c-4314-8044-99ac1a99a3ad","order_by":6,"name":"Yeju Wang","email":"","orcid":"","institution":"HanZhong Central Hospital","correspondingAuthor":false,"prefix":"","firstName":"Yeju","middleName":"","lastName":"Wang","suffix":""},{"id":304871583,"identity":"c4077db1-b888-42a2-8b72-1f53d3eb32a0","order_by":7,"name":"Yijia Li","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA0klEQVRIie3RsQrCMBCA4YRAXAJZU5D6CoGC+jgJQjehYweRgFIHLV3jW+jmZl06xb1jfYOKIA4dLDgqNm4O+ef7OI4DwOX6wyiDCRCcEZhl10rEs27iaVSAOhr7SOcBr0zRTXiJQ6jrOOgpMfQuS2RDyBARzmSq8jCWCgO6Wgs7sl2oopSHPmDmvLMje3VKSmkw4GzaRegdNS055hMcyQTZkNeWAKoQAyvi6RfxoTaICVOQzlvowIxupJm3r9zA6yOe+XSVfidvkd/GXS6Xy/WxJwivSX4GgZQUAAAAAElFTkSuQmCC","orcid":"","institution":"HanZhong Central Hospital","correspondingAuthor":true,"prefix":"","firstName":"Yijia","middleName":"","lastName":"Li","suffix":""}],"badges":[],"createdAt":"2024-05-12 15:12:45","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-4408975/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-4408975/v1","draftVersion":[],"editorialEvents":[{"content":"https://doi.org/10.1038/s41598-025-86819-w","type":"published","date":"2025-02-04T15:57:25+00:00"}],"editorialNote":"","failedWorkflow":false,"files":[{"id":57517105,"identity":"5c4c681b-47e6-4d8a-bc8b-5907c281b632","added_by":"auto","created_at":"2024-05-31 20:11:45","extension":"jpg","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":47586,"visible":true,"origin":"","legend":"\u003cp\u003eWorkflow diagram of this study.\u003c/p\u003e","description":"","filename":"Figuer1.jpg","url":"https://assets-eu.researchsquare.com/files/rs-4408975/v1/4a131d509829afd87c1d8d9b.jpg"},{"id":57517102,"identity":"1ec9a11d-3660-429c-808f-3e47e90cb22c","added_by":"auto","created_at":"2024-05-31 20:11:45","extension":"jpg","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":45934,"visible":true,"origin":"","legend":"\u003cp\u003eVarious performance evaluations of the ThyroNet-X4 Genesis model. Among them, 4a is the pre training iteration diagram of 100 epoch, 4b is the F1 training iteration diagram of 100 epoch, 4c is the ACC training iteration diagram of 100 epoch, and 4d is the confusion matrix diagram.\u003c/p\u003e","description":"","filename":"Figure2.jpg","url":"https://assets-eu.researchsquare.com/files/rs-4408975/v1/679fca3d10f0782600aef6c0.jpg"},{"id":57517328,"identity":"aa45ad0f-fa66-4235-9ca4-8f9ce12c8fcd","added_by":"auto","created_at":"2024-05-31 20:19:45","extension":"jpg","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":46357,"visible":true,"origin":"","legend":"\u003cp\u003eConfusion matrix of each comparison model, where 5a represents the Baseline model; 5b represents the Baseline+Bottleneck model; 5c represents the Baseline+Block model; 5d is the VGGNet model; 5e is the AlexNet model; 5f is the GoogleNet model; 5g is the MobileNet model.\u003c/p\u003e","description":"","filename":"Figure3.jpg","url":"https://assets-eu.researchsquare.com/files/rs-4408975/v1/9c2fd2ef10d2ad55ba95e71c.jpg"},{"id":57517961,"identity":"9b9ae29f-8986-47bf-95e7-dac6afe5c56f","added_by":"auto","created_at":"2024-05-31 20:27:45","extension":"jpg","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":64269,"visible":true,"origin":"","legend":"\u003cp\u003eGrad-CAM visual activation heat map of the deep learning model.\u003c/p\u003e","description":"","filename":"Figure4.jpg","url":"https://assets-eu.researchsquare.com/files/rs-4408975/v1/16bcce2de76f9e4827e9b7f6.jpg"},{"id":57517106,"identity":"fe0e2416-08f9-43ff-9c6d-ce0c9029193b","added_by":"auto","created_at":"2024-05-31 20:11:45","extension":"jpg","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":59438,"visible":true,"origin":"","legend":"\u003cp\u003eDisplay of the data set. Among them, 2a means Malignant and 2b means benign.\u003c/p\u003e","description":"","filename":"Figure5.jpg","url":"https://assets-eu.researchsquare.com/files/rs-4408975/v1/35c0c9e10810f99c9a822b78.jpg"},{"id":57517107,"identity":"04a4ad07-0605-4a40-bc86-20abce74ffa7","added_by":"auto","created_at":"2024-05-31 20:11:45","extension":"jpg","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":54934,"visible":true,"origin":"","legend":"\u003cp\u003eThyroNet-X4 Genesis model architecture diagram.\u003c/p\u003e","description":"","filename":"Figure6.jpg","url":"https://assets-eu.researchsquare.com/files/rs-4408975/v1/250db8a812945378fee0a2f8.jpg"},{"id":75931334,"identity":"e485793a-76d7-4255-b95e-daf694273cfc","added_by":"auto","created_at":"2025-02-10 16:14:35","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1008930,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-4408975/v1/b18a5197-b5d5-46da-9c11-81adfc12d257.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"\u003cp\u003eThyroNet-X4 Genesis: An Advanced Deep Learning Model for Auxiliary Diagnosis of Thyroid Nodules' Malignancy\u003c/p\u003e","fulltext":[{"header":"1. Introduction","content":"\u003cp\u003eThyroid nodules are defined as space-occupying lesions within the thyroid gland that can be detected through imaging studies, differentiated from the surrounding thyroid tissue. These nodules can be benign or malignant, and according to recent epidemiological studies, up to 60% of the population have thyroid nodules[\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e], with a malignancy rate of approximately 1%-5%[\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e]. Further research indicates that the prevalence of thyroid nodules over 0.5 cm in diameter is 20.43%, with 8%-16% being malignant[\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e]. The treatment approaches for thyroid nodules vary based on their nature; benign nodules often require no treatment but regular follow-up. However, surgical intervention becomes necessary when nodules grow large enough to cause compressive symptoms such as difficulty breathing, swallowing difficulties, and hoarseness. Malignant thyroid nodules, posing a threat to the patient's life and quality of life, require accurate diagnosis and surgical treatment, making the differentiation of their nature a core aspect of assessment.\u003c/p\u003e \u003cp\u003eIn the general population, the incidence of palpably detected thyroid nodules is between 3% and 7%, but with the aid of high-resolution ultrasound, detection rates can soar to between 20% and 76%[\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e]. Compared to other diagnostic modalities like X-rays, MRI, and CT, ultrasound offers advantages such as efficiency, convenience, and the absence of radiation. With advancements in ultrasound resolution, technologies such as ultrasound contrast enhancement and elastography have rapidly evolved, making color Doppler ultrasound the preferred method for diagnosing thyroid nodules[\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e]. In 2011, Russ and colleagues[\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e] used indicators such as very low echogenicity, microcalcifications, an aspect ratio\u0026thinsp;\u0026gt;\u0026thinsp;1, and irregular margins or borders to develop a five-tier thyroid imaging reporting and data system (TIRADS), assessing the malignancy risk of thyroid nodules and facilitating the identification and further management of potentially malignant nodules. Subsequently, South Korea, Europe, the United States, and China have progressively established their own TIRADS[\u003cspan additionalcitationids=\"CR8 CR9\" citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e], which similarly use solid composition, low echogenicity, irregular margins, vertical growth, and microcalcifications as indicators for suspicious malignancy in assessing and grading the risk of thyroid nodules. However, the process of describing ultrasound characteristics of thyroid nodules and quantifying risk levels according to TIRADS standards can be time-consuming and may vary in accuracy due to the experience of the ultrasound physicians and the quality of the diagnostic equipment.\u003c/p\u003e \u003cp\u003eWith the rise of artificial intelligence, more studies are employing deep learning for ultrasound detection of thyroid nodules. Chi[\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e] and others proposed the GoogLeNet model, extracting features from thyroid ultrasound images and inputting these into a random forest classifier to distinguish between benign and malignant thyroid nodules. Wang[\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e] and others improved the Faster RCNN model to better extract ultrasound features of thyroid papillary carcinoma, enhancing diagnostic accuracy. Liang[\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e] developed a deep learning model specifically for classifying thyroid and breast nodules. Zhang[\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e] utilized the YOLOv3 model to discriminate between benign and malignant thyroid nodules in TIRADS category 4, significantly impacting subsequent treatment decisions and patient outcomes. Moussa[\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e] used the ImageNet-pretrained ResNet50 for transfer learning, achieving promising diagnostic results in their own ultrasound image dataset. Kwon[\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e] employed a pretrained VGG16 model for transfer learning, effectively classifying thyroid nodules based on malignancy. Clearly, deep learning holds significant value in enhancing the accuracy of thyroid nodule diagnoses, reducing physician workload, and standardizing diagnostic procedures. However, most current studies focus on single-modality ultrasound images or are only internally validated on a single dataset, with further improvements needed in accuracy, model generalization, and stability.\u003c/p\u003e \u003cp\u003eIn this study, we innovatively propose the CNN-based ThyroNet-X4 Genesis model, initially trained and cross-validated using publicly available database data, while also collecting ultrasound imaging and related clinical data from our center as an external validation set to assess the model's generalization ability and practical value. By increasing the expansion factor and the size of the convolutional kernels without adding computational burden, we significantly enhanced the model's expressive capacity and its ability to extract ultrasound features. Additionally, we incorporated grouped convolutions as an effective method to reduce the number of parameters and computational complexity, enhancing the model's learning ability across different feature channel groups. The ThyroNet-X4 Genesis model achieved optimal balance across all configurations, exhibiting the lowest training losses and the highest training and validation accuracies, effectively boosting the model's generalization capabilities and demonstrating the importance of innovative network design ideas in enhancing the performance of deep learning models. The workflow of this work is shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e"},{"header":"2. Results","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003e2.1 Internal Validation Results\u003c/h2\u003e \u003cdiv id=\"Sec4\" class=\"Section3\"\u003e \u003ch2\u003e2.1.1 ThyroNet-X4 Genesis Model Results\u003c/h2\u003e \u003cp\u003eIn our study, the ThyroNet-X4 Genesis model demonstrated excellent performance in internal validation, indicating its efficiency and accuracy in the task of differentiating benign and malignant thyroid nodules. We utilized the Adam optimizer with an initial learning rate set to 0.001 and implemented a learning rate decay strategy to finely tune the training process. The batch size was adjusted to 32 based on experimental settings and hardware configuration, ensuring sufficient data loading without excessive resource consumption. During the validation phase, the ThyroNet-X4 Genesis model exhibited superior generalization capability. Figure\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e illustrates the relevant results over 100 training epochs.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eTable\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e presents the specific training and validation results of the model. In this table, the ThyroNet-X4 Genesis model achieved excellent training loss and training accuracy across all configurations. Additionally, it obtained relatively high accuracy on the validation set, indicating that the model not only learns effectively but also possesses good predictive ability on unseen data.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eResults of ThyroNet-X4 Genesis model\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"5\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eModel\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003etrain_loss\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003etrain_accuracy\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eval_loss\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eval_accuracy\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eThyroNet-X4 Genesis(best)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.2595\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.8555\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.6149\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.7170\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec5\" class=\"Section3\"\u003e \u003ch2\u003e2.1.2 Comparative Model Results\u003c/h2\u003e \u003cp\u003eIn this study, through a series of ablation experiments, we extensively explored the performance differences among various model variants to assess the impact of structural adjustments (as shown in Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e and Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e). The baseline model exhibited a relatively high training accuracy (84.926%) but a relatively lower validation accuracy (69.5418%), implying possible overfitting. The improved Block variant (Baseline\u0026thinsp;+\u0026thinsp;Block) demonstrated better generalization capability with slightly lower training accuracy (84.7028%) and an increased validation accuracy of 70.6199%, suggesting that the additional Block structure helps improve the model's performance on unseen data. Although the Bottleneck variant (Baseline\u0026thinsp;+\u0026thinsp;Bottleneck) achieved the highest training accuracy (85.237%), the validation accuracy decreased to 68.4636%, indicating exacerbated overfitting.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003ePerformance of each comparison model\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"5\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eModel\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003etrain_loss\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003etrain_accuracy\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eval_loss\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eval_accuracy\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eBaseline\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.2795\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.8493\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e1.393\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.6954\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eBaseline\u0026thinsp;+\u0026thinsp;Bottleneck\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.2740\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.8524\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e1.1411\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.6846\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eBaseline\u0026thinsp;+\u0026thinsp;Block\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.2751\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.8470\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.6972\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.7061\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eVGGNet\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.6094\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.6694\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e1.1586\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.5983\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAlexNet\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.5822\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.6561\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e2.2945\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.4952\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGoogleNet\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.9523\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.6728\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e1.6618\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.5705\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMobileNet\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.5803\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.6885\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e1.6088\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.5971\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eThe ThyroNet-X4 Genesis model showed the optimal balance across all configurations, with the lowest training loss (0.259516), the highest training (85.5478%), and validation accuracy (71.6981%), significantly enhancing its generalization capability. Additionally, comparisons were made with other deep learning architectures such as VGGNet, AlexNet, GoogleNet, and MobileNet. VGGNet performed well in extracting shape and texture features due to its deep but simple convolutional structure, yet it exhibited the lowest validation accuracy (59.83%), suggesting its potential inadequacy in handling the complexity of thyroid nodules. AlexNet and MobileNet showed relatively lower validation accuracy (49.52% and 59.71%, respectively). Although they are efficient in resource-constrained environments, they have limitations in handling fine-grained features. GoogleNet, with its Inception module effectively capturing features at different scales, still had a low validation accuracy (57.05%), indicating that its complex structure might be overly intricate for the dataset used in this study.\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv id=\"Sec6\" class=\"Section2\"\u003e \u003ch2\u003e2.2 External Center Validation and Clinical Interpretability\u003c/h2\u003e \u003cp\u003eTo further evaluate the potential clinical application of the ThyroNet-X4 Genesis model, we applied it to ultrasound images of 658 thyroid nodule patients from our center (with a test accuracy of 67.02%) and conducted a detailed analysis of its clinical interpretability. We employed Grad-CAM (Gradient-weighted Class Activation Mapping) technology to generate heatmaps, highlighting the regions of interest the model focused on (as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e). This visualization technique helps doctors verify the model's rationale and ensure that its focus aligns with the clinical diagnostic process.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eTo further assess the model's clinical interpretability, we invited three ultrasound doctors with years of thyroid diagnosis experience to independently evaluate the model's diagnostic results. The doctors assessed the model's diagnostic results based on their professional knowledge and pointed out whether the features the model focused on in certain cases aligned with clinical diagnostic criteria. The results showed that in most cases, the regions of interest identified by ThyroNet-X4 Genesis were consistent with those identified by the doctors, and the model's diagnostic accuracy for malignant nodules was comparable to that of the doctors.\u003c/p\u003e \u003cp\u003eFurthermore, we conducted in-depth analysis of misdiagnosed and missed cases by the model to assess its potential clinical risks. We found that in certain types of thyroid nodules, such as mixed nodules and ectopic thyroid nodules, the model's diagnostic accuracy was relatively low. These types of nodules often exhibit complex morphological features in ultrasound images, making it easy for the model to confuse benign and malignant conditions. In such cases, the model's diagnostic results can serve as supplementary information for doctors rather than the sole basis for diagnosis.\u003c/p\u003e \u003c/div\u003e"},{"header":"3. Discussion","content":"\u003cp\u003eIn this study, we proposed a novel deep learning model, ThyroNet-X4 Genesis, which trained on thyroid ultrasound images from publicly available medical imaging databases to classify the benign and malignant nature of thyroid nodules. We used ultrasound images from our center as the validation set, and the results showed that the ThyroNet-X4 Genesis model exhibited the lowest training loss (0.259516), highest training accuracy (85.5478%), and highest validation accuracy (71.6981%) across all configurations, demonstrating its superior generalization capability.\u003c/p\u003e \u003cp\u003eSeveral reasons contribute to the optimal balance demonstrated by this model. Firstly, the expansion coefficient was increased from 1 to 4, significantly enlarging the channel number of the network's output layer. This improvement allowed the network to carry more information without significantly increasing computational burden, thus significantly enhancing the model's expressive power. Secondly, improvements were made in convolution kernel size and grouped convolution. We used a 5x5 convolution kernel in the model's second convolutional layer to capture broader contextual information and enhance feature extraction capabilities. Additionally, introducing grouped convolution as an effective method to reduce parameter count and computational complexity strengthened the model's ability to learn from different feature channel groups. Lastly, the model's training set sourced from publicly available medical imaging databases, indicating that the model learned from ultrasound images from different medical centers, thus enhancing its generalization.\u003c/p\u003e \u003cp\u003eThe core of thyroid nodule assessment lies in distinguishing between benign and malignant nodules. Thyroid ultrasound is the preferred examination for assessing nodule malignancy risk based on ultrasound image features and graded according to the TIRADS criteria. However, this process relies on the experience of ultrasound doctors and ultrasound equipment conditions, leading to a certain degree of misdiagnosis. In recent years, with the continuous advancement of deep learning technology, an increasing number of models have been developed for the diagnosis of thyroid nodule malignancy, improving the efficiency and accuracy of thyroid nodule diagnosis. For instance, Guan et al. [\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e] used the InceptionV3 model to classify thyroid ultrasound images, achieving high sensitivity and specificity in the test group. Ma et al. [\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e] fused two CNN-based models for classifying thyroid nodule malignancy, achieving an internal accuracy of up to 83.02%. PENG et al. [\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e] developed the ThyNet model, which showed significantly higher diagnostic accuracy than professional doctors. However, most studies to date have not applied the models to external data for validation, limiting the models' universality and clinical application value. In this study, the ThyroNet-X4 Genesis model achieved a training accuracy of up to 85.6%, validation accuracy of 71.7%, and external validation accuracy of 67.0%, indicating its potential for widespread application.\u003c/p\u003e \u003cp\u003eHowever, this study also has some limitations. Firstly, the ThyroNet-X4 Genesis model proposed in this study only identifies the benign and malignant types of thyroid nodules, without classifying benign lesions such as inflammation, cysts, and adenomatous nodules. The included malignant thyroid nodule pathology results were all thyroid papillary carcinomas, and rare types of thyroid cancer such as medullary thyroid carcinoma and thyroid follicular carcinoma were not included. Future research can further classify these types. Additionally, the model was only validated at our center, and the provided malignant thyroid nodule cases were relatively few, thus the validation accuracy may be underestimated. Lastly, including data from more centers to increase the size of the training set could further improve the diagnostic accuracy of the model.\u003c/p\u003e \u003cp\u003eIn conclusion, our proposed ThyroNet-X4 Genesis model, which incorporates an increase in expansion coefficient and optimized convolutional layer configuration, has successfully improved the performance of the model in the task of benign and malignant diagnosis of thyroid nodules. These improvements not only enhanced the model's expressive power but also increased its generalization capability, providing valuable insights for the application of deep learning models in medical image analysis. Through comprehensive analysis, our best model demonstrated significant advantages in balancing computational efficiency and diagnostic accuracy, proving the importance of innovative network design concepts in enhancing the performance of deep learning models.\u003c/p\u003e"},{"header":"4. Methods","content":"\u003cdiv id=\"Sec9\" class=\"Section2\"\u003e \u003ch2\u003e4.1 Data Acquisition\u003c/h2\u003e \u003cdiv id=\"Sec10\" class=\"Section3\"\u003e \u003ch2\u003e4.1.1 Inclusion and Exclusion Criteria\u003c/h2\u003e \u003cp\u003eThe inclusion and exclusion criteria for this study are as follows. Inclusion criteria include: 1. Thyroid ultrasound assessments conducted prior to surgery or biopsy, 2. Definitive histopathological results obtained after surgery or biopsy, 3. Ultrasound images of thyroid nodules that include complete transverse and longitudinal sectional views. Exclusion criteria include: 1. Indeterminate pathological results, 2. Ultrasound images that do not display the entire extent of the nodules, patients with a history of invasive treatments such as surgery or ablation, 3. Ultrasound images that are unclear or obscured by ultrasound marker lines, blood flow signals, etc.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec11\" class=\"Section3\"\u003e \u003ch2\u003e4.1.2 Data Collection\u003c/h2\u003e \u003cp\u003eIn this study, data collection was conducted in two parts. Firstly, we utilized the publicly accessible medical imaging database DDTI (Digital Database of Thyroid Ultrasound Images), supported by the National University of Colombia, CIM@LAB, and IDIME (Institute of Medical Diagnostics). This database currently includes 299 cases and has been expanded to contain 620 benign and 914 malignant thyroid nodules, totaling over 1500 ultrasound images. Each case is presented as an XML file containing expert annotations and patient information. The database is regularly updated with new cases and images for the development of computer-aided diagnostic systems and serves as a training and teaching tool for new radiologists. These data are used for the initial training and cross-validation of the model.\u003c/p\u003e \u003cp\u003eSecondly, we collected thyroid ultrasound images and related clinical data from Hanzhong Central Hospital in Shaanxi Province, conducting a retrospective analysis of data from 155 patients treated at our center, including 55 benign nodules and 100 malignant nodules, totaling 658 ultrasound images (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e). This dataset did not participate in the initial model training but served as an external validation set to assess the model's generalization ability and practical application value. It is noteworthy that this retrospective study was conducted in accordance with the Declaration of Helsinki and was approved by the Ethics Committee of Hanzhong Central Hospital (Approval No.2024(18)), and waived the requirement for the written informed consent of the patients, because the selected clinical and imaging data in this retrospective study would not affect the prognosis and privacy of the patients.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv id=\"Sec12\" class=\"Section2\"\u003e \u003ch2\u003e4.2 Data Preprocessing\u003c/h2\u003e \u003cp\u003eThe collected images underwent preprocessing, which included steps such as denoising, normalization, and resizing to a uniform size to ensure the quality and consistency of the image data input into the model. Additionally, techniques such as rotation, scaling, and mirroring were employed to enhance the diversity of the image data and improve the model's generalization ability.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec13\" class=\"Section2\"\u003e \u003ch2\u003e4.3 Construction of Diagnostic Model\u003c/h2\u003e \u003cdiv id=\"Sec14\" class=\"Section3\"\u003e \u003ch2\u003e4.3.1 Proposal of a Deep Learning-Based Model\u003c/h2\u003e \u003cp\u003eDeep learning models facilitate the extraction and classification of benign and malignant characteristics from ultrasound images of thyroid nodules. U-NET-based models are widely used in the diagnosis of thyroid nodules; Wu[\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e] and others built on the U-NET to introduce a method for ultrasound image segmentation of thyroid nodules based on joint upsampling, achieving precise localization of the target thyroid. However, this model is more complex than the U-NET, resulting in longer computation times. The ReAgU-Net model proposed by Ding[\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e] increases the backward propagation gradient to address the loss of spatial information due to increased network depth, although its performance diminishes when the contrast between nodules and background is low. To enhance the accuracy of AI diagnostics of thyroid nodules, Wei[\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e] and others improved upon DenseNet to develop a precise post-localization integrated deep learning classification model for thyroid nodules, though this model does not analyze a wide range of thyroid nodule pathologies and only provides classification results without standards or texture analysis. In this study, we innovatively propose the ResNet-based ThyroNet-X4 Genesis model.\u003c/p\u003e \u003cp\u003eThe ThyroNet-X4 Genesis model developed in this research is a deep convolutional neural network (CNN) specifically designed for the automatic identification and classification of the benign and malignant nature of thyroid nodules. The model integrates multiple layers of convolutional and pooling layers, utilizing 3x3 and 5x5 convolutional kernels to intricately capture the shape, edges, and texture of the nodules. To enhance the learning efficiency of the deep network and prevent issues of gradient vanishing, the model incorporates techniques from residual networks (ResNet) and densely connected networks (DenseNet), which bolster feature transmission and reuse through residual and dense connections, respectively. Moreover, an expansion coefficient was introduced before the output layer, increasing from 1 to 4, which widens the network, enabling it to handle more information without significantly increasing the computational burden. The use of grouped convolution techniques also helps to reduce the number of parameters and computational complexity. These innovative designs have resulted in exceptional performance of the ThyroNet-X4 Genesis during initial training and cross-validation on public medical imaging databases, as well as demonstrating outstanding generalization capability and high accuracy on the external validation set at our center. Specific architectural and parameter details of the model can be found in Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003e (Model Architecture Diagram).\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eThe ThyroNet-X4 Genesis model is equipped with a multi-layer convolutional network structure that integrates residual and dense connectivity technologies, as well as optimized computational efficiency through expansion coefficients and grouped convolution. These features enable it to excel in the diagnosis of benign and malignant thyroid nodules. Similarly, these characteristics are applicable to the task of identifying breast tumors, as the diagnosis of breast tumors requires in-depth analysis of complex image features such as irregular margins and echo heterogeneity, which are also common in thyroid nodule images. The advanced feature processing capabilities and excellent generalization ability of the ThyroNet-X4 Genesis model demonstrate its effective adaptability for the recognition and classification of breast tumors, showcasing its potential for broad application in the field of medical image analysis.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec15\" class=\"Section3\"\u003e \u003ch2\u003e4.3.2 Comparative Models\u003c/h2\u003e \u003cp\u003eIn our study, the performance of the ThyroNet-X4 Genesis model was thoroughly evaluated by comparing it against a range of deep learning architectures. These included the traditional ResNet34, which served as a baseline model and has demonstrated excellent performance in medical image processing due to its residual network architecture. Additionally, the enhanced Baseline\u0026thinsp;+\u0026thinsp;Block variant aimed to improve diagnostic accuracy by strengthening local feature processing, while the Baseline\u0026thinsp;+\u0026thinsp;Bottle variant introduced bottleneck layers to optimize the use of computational resources and speed up processing. VGGNet is known for its deep, simple convolutional structure that excels in extracting image textures and shapes. Although simpler in structure, AlexNet is notable for its effectiveness in rapid preliminary feature extraction. GoogleNet, with its Inception modules, captures image features effectively at various scales, making it well-suited for handling complex thyroid nodule image data. MobileNet demonstrates efficient performance under resource-constrained conditions, especially suitable for quickly processing large volumes of data. These comparisons not only highlighted the superior performance of ThyroNet-X4 Genesis in diagnosing thyroid nodules but also emphasized its efficiency and potential in handling complex medical image data, proving its effectiveness and innovativeness as a diagnostic tool for thyroid nodules.\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv id=\"Sec16\" class=\"Section2\"\u003e \u003ch2\u003e4.4 Experimental Setup\u003c/h2\u003e \u003cp\u003eIn this study, we acquired thyroid ultrasound image data from a public database and precisely segmented it into training and validation sets. Specifically, after image processing and data augmentation, the training set included 622 malignant and 624 benign images, while the validation set comprised 292 malignant and 286 benign images. Additionally, data collected from Hanzhong Central Hospital served as an external test set to further validate the model's generalization ability and practical application effectiveness, containing 192 benign and 410 malignant thyroid ultrasound images.\u003c/p\u003e \u003cp\u003eThe experiments were conducted on an Ubuntu 20.04 operating system, programmed using Python 3.8, and primarily utilizing PyTorch 1.10.0 as the deep learning framework, with computational acceleration provided by CUDA 11.3. In terms of hardware, our laboratory was equipped with RTX A5000 GPUs and a server powered by an Intel(R) Xeon(R) Platinum 8358P CPU @ 2.60GHz with 15 vCPUs, ensuring ample computational resources and efficient processing capabilities.\u003c/p\u003e \u003cp\u003eDuring the model training phase, we rigorously optimized all model parameters, including adjustments to learning rates and batch sizes, to ensure optimal performance on the training and validation sets and to prevent overfitting. Our tuning methods ensured the stability and reliability of the models, while independent external test sets were used to evaluate performance on unseen data, verifying their accuracy and applicability in real-world applications.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec17\" class=\"Section2\"\u003e \u003ch2\u003e4.5 Model Evalution\u003c/h2\u003e \u003cp\u003eWhen evaluating deep learning models, we often use several key metrics to measure performance, including accuracy (ACC), F1 score, and so on. These evaluation metrics collectively describe the model's performance in various aspects, including prediction accuracy, comprehensiveness, and consistency between predicted and actual results. First, accuracy (ACC) is the most intuitive evaluation metric, representing the ratio of correctly classified sample data to the total number of samples. Its mathematical formula is expressed as follows:\u003cdiv id=\"Equ1\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ1\" name=\"EquationSource\"\u003e\n$$ACC=\\frac{{(TP+TN)}}{{(TP+TN+FP+FN)}}$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e1\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eWhere TP represents the number of true positive samples, TN represents the number of true negative samples, FP represents the number of false positive samples, and FN represents the number of false negative samples. A higher accuracy indicates a more effective classifier and higher precision in the predicted results.\u003c/p\u003e \u003cp\u003eSecondly, the F1 score is the harmonic mean of precision and recall. Precision represents the number of samples determined as positive examples, while recall represents the proportion of correctly predicted positive samples out of all actual positive samples. The formula for calculating the F1 score is:\u003cdiv id=\"Equ2\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ2\" name=\"EquationSource\"\u003e\n$$PRE=\\frac{{TP}}{{(TP+FP)}}$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e2\u003c/div\u003e\u003c/div\u003e\u003cdiv id=\"Equ3\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ3\" name=\"EquationSource\"\u003e\n$$REC=\\frac{{TP}}{{(TP+FN)}}$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e3\u003c/div\u003e\u003c/div\u003e\u003cdiv id=\"Equ4\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ4\" name=\"EquationSource\"\u003e\n$${F_1}=\\frac{{2P * R}}{{P+R}}$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e4\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003c/div\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eFunding\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis article has no funding.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAcknowledgements\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eWe thank National University of Colombia, CIM@LAB, and IDIME for the open data resources and all investigators who participated in the study.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthor contributions statement\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eX.X.W and Y.P.N were responsible for the conceptualization, methodology, software, investigation, formal analysis, and writing - original draft.their contributions to this study are the same, so X.X.W and Y.P.N are the co-first authors.\u003c/p\u003e\n\u003cp\u003eL.L was responsible for conceptualization, methodology, data collection, writing-original draft.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eT was responsible for methodology, software, investigation.\u003c/p\u003e\n\u003cp\u003eZ, Y.M.W and Y.J.W were responsible for data collection.\u003c/p\u003e\n\u003cp\u003eJ.L was responsible for the design and review of the study, she is the co-corresponding authors of this study.\u003c/p\u003e\n\u003cp\u003eAll authors reviewed the manuscript.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eData availability statement\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe datasets, from Hanzhong Central Hospital in Shaanxi Province, China, analyzed and generated in this study are not publicly available because the institution prohibits the public upload of any patient\u0026apos;s private data. However, partial datasets are available from the corresponding author upon reasonable request, please contact [email protected] by email.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAdditional Information\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThere are no conflict of interest in this study. All authors have read and approved this version of the article ,and due care has been taken to ensure the integrity of the work . Neither the entire paper nor any part of its content has been published or has been accepted elsewhere . It is not being submitted to any other journal.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eEthics approval:\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eEthical Committee: The Biomedical Ethics Review Committee of\u0026nbsp;HanZhong Central Hospital.\u003c/p\u003e\n\u003cp\u003eEthical Approval Number:No.2024（18）.\u003c/p\u003e\n\u003cp\u003eThis was a retrospective study, and the study data were partly derived from the publicly accessible medical imaging database DDTI (Digital Database of Thyroid Ultrasound Images), supported by the National University of Colombia, CIM@LAB, and IDIME (Institute of Medical Diagnostics) and partly from Hanzhong Central Hospital in Shaanxi Province, China. The selected clinical and imaging data was waived the requirement for the written informed consent of the patients, because the selected clinical and imaging data in this retrospective study would not affect the prognosis and privacy of the patients.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthor Statement:\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthor 1\u003c/strong\u003e: Conceptualization, Methodology, Software, Investigation, Formal Analysis, Writing - Original Draft.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthor 2\u0026nbsp;\u003c/strong\u003e:\u0026nbsp;Conceptualization, Methodology, Data collection, Writing-original draft.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthor 3\u003c/strong\u003e: Methodology, Software, Investigation.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthor 4,5 and 6\u003c/strong\u003e: Data collection.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCorresponding Author:\u003c/strong\u003eDesigning and review of the study.\u003c/p\u003e\n\u003cp\u003eAll authors reviewed the manuscript.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eGuth S Theune U Aberle J Galach A \u0026amp; Bamberger CM. Very high prevalence of thyroid nodules detected by high frequency (13 MHz) ultrasound examination. European Journal of Clinical Investigation 200939699\u0026ndash;706. (10.1111/j.1365-2362.2009.02162.x)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGrussendorf M Ruschenburg I \u0026amp; Brabant G. Malignancy rates in thyroid nodules: a long-term cohort study of 17,592 patients. European Thyroid Journal 202211 e220027. (10.1530/ETJ-22-0027)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLi Y, Teng D, Ba J, et al. Efficacy and Safety of Long-Term Universal Salt Iodization on Thyroid Disorders: Epidemiological Evidence from 31 Provinces of Mainland China[ J]. Thyroid, 2020,30(4):568\u0026ndash;579. DOI: 10. 1089 / thy. 2019. 0067.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDurante C, Grani G, Lamartina L, et al. The Diagnosis and management of thyroid nodules: a review[ J]. JAMA, 2018,319 (9):914\u0026ndash;924. DOI: 10. 1001 / jama. 2018. 0898.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGharib H, Papini E, Garber JR, et al. American Association of Clinical Endocrinologists, American College of Endocrinology, and Associazione Medici Endocrinologi Medical Guidelines for Clinical Practice for the Diagnosis and Management of Thyroid Nodules\u0026mdash;2016 Update [ J]. Endocr Pract, 2016,22 (5):622\u0026ndash;639. DOI:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.4158/EP161208\u003c/span\u003e\u003cspan address=\"10.4158/EP161208\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. GL.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRuss G,Bigorgne C,Royer B,et al.The Thyroid Imaging Reporting and Data System (TIRADS) for ultrasound of the thyroid[J].J Radiol,2011,92(7/8):701\u0026ndash;713.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLee JY, Baek JH, Ha EJ, et al. 2020 Imaging Guidelines for Thyroid Nodules and Differentiated Thyroid Cancer: Korean Society of Thyroid Radiology[J]. Korean J Radiol, 2021,22(5):840\u0026ndash;860.DOI: 10. 3348 / kjr. 2020. 0578.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhou J, Yin L, Wei X, Zhang S, Song Y, Luo B, et al.. 2020 Chinese guidelines for ultrasound Malignancy risk stratification of thyroid nodules: the C-TIRADS. Endocrine. (2020) 70:256\u0026ndash;79. doi: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1007/s12020-020-02441-y\u003c/span\u003e\u003cspan address=\"10.1007/s12020-020-02441-y\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLeenhardt L, Erdogan MF, Hegedus L, et al. 2013 European thyroid association guidelines for cervical ultrasound scan and ultrasound-guided techniques in the postoperative management of patients with thyroid cancer [ J]. Eur Thyroid J, 2013, 2 (3): 147\u0026ndash;159.DOI: 10. 1159 / 000354537.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKwak JY, Han KH, Yoon JH, et al. Thyroid imaging reporting and data system for US features of nodules: a step in establishing better stratific-ation of cancer risk[ J]. Radiology, 2011,260 (3):892\u0026ndash;899. DOI: 10. 1148 / radiol. 11110206.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChi J, Walia E, Babyn P, et al. Thyroid Nodule Classification in Ultrasound Images by FineTuning Deep Convolutional Neural Network[J]. Journal of Digital Imaging, 2017, 30(4): 477\u0026ndash;486.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWang Y H, Ke W, Wan P. A Method of Ultrasonic Image Recognition for Thyroid Papillary Carcinoma Based on Deep Convolution Neural Network[J]. NeuroQuantology, 2018, 16(5).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLiang X, Yu J, Liao J, et al. Convolutional Neural Network for Breast and Thyroid Nodules Diagnosis in Ultrasound Imaging[J]. BioMed Research International, 2020, 2020:1\u0026ndash;9.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhang X, Jia C, Sun M, Ma Z. The application value of deep learning-based nomograms in benign-malignant discrimination of TI-RADS category 4 thyroid nodules. Sci Rep. 2024;14(1):7878. doi: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1038/s41598-024-58668-6\u003c/span\u003e\u003cspan address=\"10.1038/s41598-024-58668-6\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. PMID: 38570589; PMCID: PMC10991510.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMoussa O, Khachnaoui H, Guetari R, et al. Thyroid nodules classification and diagnosis in ultrasound images using fine-tuning deep convolutional neural network[J]. International Journal of Imaging Systems and Technology, 2019, 30(1):185\u0026ndash;195.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKwon S W, Choi I J, Kang J Y, et al. Ultrasonographic Thyroid Nodule Classification Using a Deep Convolutional Neural Network with Surgical Pathology[J]. Journal of digital imaging,2020.Kumar V, Webb J, Gregory A, et al. Automated segmentation of thyroid nodule, gland, and cystic components from ultrasound images using deep learning. IEEE Access, 2020, 8: 63482\u0026ndash;63496.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWu J, Zhang Z, Zhao J, et al. Ultrasound image segmentation of thyroid nodules based on joint up-sampling//The 2020 Second International Conference on Artificial Intelligence Technologies and Application (ICAITA), Dalian: AEIC-Academic Exchange Information Center, 2020, 1651(1): 012157.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDing J, Huang Z, Shi M, et al. Automatic thyroid ultrasound image segmentation based on U-shaped network//2019 12th International Congress on Image and Signal Processing, Biomedical Engineering and Informatics (CISP-BMEI), Suzhou: IEEE, 2019: 1\u0026ndash;5.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWei X, Gao M, Yu R, et al. Ensemble deep learning model for multicenter classification of thyroid nodules on ultrasound images.Medical Science Monito, 2020, 26: e926096.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGuan Q, Wang Y, Du J, et al. Deep learning based classification of ultrasound images for thyroid nodules: a large scale of pilot study.Annals of Translational Medicine, 2019, 7(7): 137.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMa J, Wu F, Zhu J, et al. A pre-trained convolutional neural network based method for thyroid nodule diagnosis. Ultrasonics. 2017;73:221\u0026ndash;30.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePENGS,LIUY,LV,etal.Deeplearning-basedartificialintelligencemodeltoassistthyroidnodulediagnosisand management:a multicentrediagnosticstudy[J].LancetDigit Health,2021,3(4):e250.\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":true,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"scientific-reports","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"scirep","sideBox":"Learn more about [Scientific Reports](http://www.nature.com/srep/)","snPcode":"","submissionUrl":"","title":"Scientific Reports","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Scientific Reports","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"Thyroid nodules, Deep learning, Auxiliary diagnosis, External validation, Clinical interpretability","lastPublishedDoi":"10.21203/rs.3.rs-4408975/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-4408975/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eThyroid nodules are common endocrine disorders, and the discrimination between benign and malignant nodules is crucial for treatment decisions. Traditional ultrasound diagnosis relies on the experience of physicians, which may pose risks of misdiagnosis. In this study, we propose a novel deep learning model, ThyroNet-X4 Genesis, for the automatic classification of thyroid nodules' malignancy. This model is based on the ResNet module, which optimizes computational efficiency and enhances feature extraction capabilities by introducing grouped convolution and increasing the convolution kernel size, thus extracting features and classifying nodules in ultrasound images. We obtained data from publicly available medical imaging databases for internal training and validation and used ultrasound images collected from HanZhong Central Hospital as an external validation set to evaluate the model's generalization ability and practical application value.ThyroNet-X4 Genesis achieved training and validation accuracies of 85.55% and 71.70%, respectively, on the internal validation set, with a testing accuracy of 67.02% on the external validation set, outperforming other mainstream comparative models, indicating its good performance in actual clinical applications. The development of this model showcases the potential of deep learning in thyroid imaging analysis, providing valuable references for future development of high-performance medical diagnostic models.\u003c/p\u003e","manuscriptTitle":"ThyroNet-X4 Genesis: An Advanced Deep Learning Model for Auxiliary Diagnosis of Thyroid Nodules' Malignancy","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2024-05-31 20:11:40","doi":"10.21203/rs.3.rs-4408975/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Revision requested","date":"2024-09-02T11:12:54+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2024-08-23T08:19:56+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2024-08-11T15:11:56+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"234470559150537232365281875784546262638","date":"2024-08-05T07:46:41+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"118141364099904689796083681826835162721","date":"2024-07-31T15:27:14+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2024-06-19T17:54:41+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2024-06-19T17:50:51+00:00","index":"","fulltext":""},{"type":"editorInvited","content":"","date":"2024-05-18T07:22:37+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2024-05-16T03:30:08+00:00","index":"","fulltext":""},{"type":"submitted","content":"Scientific Reports","date":"2024-05-12T15:11:25+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"scientific-reports","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"scirep","sideBox":"Learn more about [Scientific Reports](http://www.nature.com/srep/)","snPcode":"","submissionUrl":"","title":"Scientific Reports","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Scientific Reports","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"d473c17d-4529-477f-9fc7-8abd214bc425","owner":[],"postedDate":"May 31st, 2024","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"published-in-journal","subjectAreas":[{"id":32378575,"name":"Health sciences/Endocrinology"},{"id":32378576,"name":"Health sciences/Endocrinology/Endocrine system and metabolic diseases"},{"id":32378577,"name":"Health sciences/Endocrinology/Endocrine system and metabolic diseases/Thyroid diseases"}],"tags":[],"updatedAt":"2025-02-10T16:10:41+00:00","versionOfRecord":{"articleIdentity":"rs-4408975","link":"https://doi.org/10.1038/s41598-025-86819-w","journal":{"identity":"scientific-reports","isVorOnly":false,"title":"Scientific Reports"},"publishedOn":"2025-02-04 15:57:25","publishedOnDateReadable":"February 4th, 2025"},"versionCreatedAt":"2024-05-31 20:11:40","video":"","vorDoi":"10.1038/s41598-025-86819-w","vorDoiUrl":"https://doi.org/10.1038/s41598-025-86819-w","workflowStages":[]},"version":"v1","identity":"rs-4408975","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-4408975","identity":"rs-4408975","version":["v1"]},"buildId":"qtupq5eGEP_6zYnWcrvyt","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}