Unveiling Hidden Carcinomas: nonenhanced CT-Based Radiomics Model Enhances PTC Detection in Hashimoto's Thyroiditis Running title: CT radiomics identifies carcinoma in HT

doi:10.21203/rs.3.rs-4862440/v1

Download PDF

Research Article

Unveiling Hidden Carcinomas: nonenhanced CT-Based Radiomics Model Enhances PTC Detection in Hashimoto's Thyroiditis Running title: CT radiomics identifies carcinoma in HT

https://doi.org/10.21203/rs.3.rs-4862440/v1

This work is licensed under a CC BY 4.0 License

You are reading this latest preprint version

Background: Hashimoto's thyroiditis (HT) is a common benign thyroid disease that often coexists with papillary thyroid carcinoma (PTC). Owing to the diffuse changes in the thyroid caused by HT, PTCs can be challenging to detect using conventional imaging modalities such as ultrasound and CT. The aim of this study is to develop a radiomics model that uses nonenhancedCT (NECT) to predict the presence of PTC in the patients with HT, thereby improving early diagnostic accuracy.

Materials and Methods: This retrospective study included pathologically confirmed HT patients with or without PTC who underwent NECT scans within 30 days before surgery from January 2017 to April 2023 at the Second Affiliated Hospital of Nanchang University (Hospital I) or Ganzhou People's Hospital (Hospital II). Radiomic features were extracted using PyRadiomics. Interclass correlation coefficient, Pearson correlation and LASSO analyses were conducted to reduce the dimensionality of the radiomicfeatures. Five machine learning algorithms, including logistic regression, naive Bayes, support vector machine, k-nearest neighbor, and multilayer perceptron (MLP) classifiers, were employed to develop and validate the prediction models based on the remaining features.

Results: A total of 130 patients, 89 from Hospital I and 41 from Hospital II, were included. Six features with nonzero coefficients were retained by the LASSO algorithm for inclusion in the machine learning models. The MLP model performed the best in the external validation cohort, with an area under the curve of 0.783, a sensitivity of 64.29%, and a specificity of 92.31%.

Conclusion: A radiomics model based on NECT can identify PTCs in patients with HT and has the potential to enhance early diagnosis and intervention for these patients.

Hashimoto's thyroiditis

papillary thyroid carcinoma

radiomics

nonenhanced CT

machine learning

Hashimoto's thyroiditis (HT) is one of the most common benign thyroid disorders and is characterized by chronic lymphocytic infiltration and gradual destruction of thyroid tissue, often leading to hypothyroidism¹. Studies have shown that HT is associated with and sometimes coexists with papillary thyroid carcinoma (PTC). Indeed, the incidence of PTC is greater in HT patients than in the general population² and 19% of PTC patients have been diagnosed with HT^3,4. The underlying mechanisms of this association are multifactorial and involve chronic inflammation, autoimmune processes, and potential genetic predispositions^5–7. Guidelines dictate that HT patients usually need only regular follow-up or medication, and surgery is not required unless there is a concomitant malignancy or symptoms of compression¹. Identifying PTCs within the inflammatory milieu of HTs is crucial for decision making to support timely and effective management.

Currently, the gold standard for the diagnosis of PTC is still histopathological examination based on invasive fine-needle aspiration biopsy; however, biopsy sampling errors can lead to missed diagnosis of PTC⁸. In contrast, noninvasive imaging methods, such as ultrasound and computed tomography (CT), although widely used, have sensitivity and for this application. HT typically presents as bilateral diffuse enlargement of the thyroid with heterogeneously decreased echo density⁹, while PTCs commonly appear as nodules. When PTCs, especially micro-PTCs, coexist with HT, they are easily masked by diffuse thyroid changes caused by HT. Therefore, more advanced techniques and methods are needed to reduce missed detection of PTC.

Radiomics refers to the high-throughput extraction of quantitative features from medical images and subsequent construction of machine learning (ML) models with extracted features to predict genomic patterns and clinical outcomes. It has been widely applied in medical imaging analysis, especially in oncology^10,11. For example, Fang et al. suggested that ultrasound radiomics can distinguish between benign and malignant thyroid nodules in HT, confirming its role in analyzing the heterogeneity of thyroid lesions¹². The purpose of this study was to develop a radiomics model to predict the presence of PTC against the background of HT from nonenhanced CT (NECT).

Patients

This retrospective study was approved by the research ethics board, which waived the requirement for informed consent. Patients with pathologically confirmed HT treated at the Second Affiliated Hospital of Nanchang University (hospital I) or Ganzhou People's Hospital (hospital II) from January 2017 to April 2023 were included. The inclusion criteria were (1) neck NECT performed within 30 days prior to surgery and (2) histologically confirmed HT with/without PTC. The exclusion criteria included (1) a lack of high-quality CT images and (2) from the presence of other malignancies.

Finally, 130 patients were enrolled in this study; among them, 89 patients were from hospital I and 41 patients were from hospital II. (Fig. 1).

CT scanning

At hospital I, all neck NECT scans were obtained using a 128-slice machine (SOMATOM Definition Flash). For hospital II, all neck NECT scans were obtained using a 512-slice machine (Revolution CT; GE Healthcare). The parameters were as follows: 120 kV; automatic tube current; scanning layer spacing, 5 mm; slice thickness, 5 mm.

Segmentation of the region of interest and extraction of radiomic features

In this study, preoperative NECT data were exported as DICOM files. Then, they were resampled to a 1 mm × 1 mm × 5 mm voxel size using 3D slicer software (version 5.2.2, https://www.slicer.org). The resampled CT images were then imported into ITK-SNAP software (version 3.8, http://www.itksnap.org/) to delineate regions of interest (ROIs). The delineation of each ROI was performed independently by two radiologists with 1 and 4 years of work experience. The ROI was drawn by outlining the edge of the thyroid and completely covering the whole gland (Fig. 2). Finally, radiomic data were extracted using the Python package PyRadiomics (version 3.0.1); a total of 851 features were extracted for each patient.

Feature selection and model establishment

All the features were normalized to 0 ~ 1. To reduce the dimensionality of the radiomic features, we adopted the following procedure: First, the interclass correlation coefficient (ICC) was used to estimate the reproducibility of each feature, and features with ICC values ≥ 0.75 were selected for further analysis. Second, a pairwise Pearson correlation analysis was used to identify pairs of features with an absolute correlation coefficient greater than 0.9; one feature from these pairs was selected at random for inclusion in further calculations. Third, the features were screened using least absolute shrinkage and selection operator (LASSO) analysis, and the optimal log (λ) value was determined through 10-fold cross validation.

Eventually, 5 ML models, including logistic regression (LR), naive Bayes (NB), support vector machine (SVM), K nearest neighbor (KNN), and multilayer perceptron (MLP) classifier, were constructed based on the remaining features.

Statistical analysis and model performance evaluation

Continuous features are reported as the means ± SDs, and categorical features are reported as counts and percentages. The age and sex of the three cohorts were compared using one-way analysis of variance (ANOVA) and the chi-square test, respectively. The models were evaluated via receiver operating characteristic (ROC) curve analysis to determine the area under the curve (AUC). Decision curve analysis (DCA) was conducted to determine the clinical usefulness of each model. All the statistical analyses were conducted using IBM SPSS Statistics (Version 26), the Python 3 programming language and R software (Version 4.1.3, www.r-project.org). A p value < 0.05 was considered significant.

Clinical characteristics

In accordance with the inclusion criteria, 98 patients were included from hospital I, and 46 patients were included from hospital II. Among the patients from hospital I, 1 patient was excluded because of the presence of other malignancies, and 9 patients were excluded because of a lack of high-quality images. Five patients from hospital II were excluded because of a lack of high-quality images. Thus, 130 patients were ultimately included in the current study, with 89 from hospital I and 41 from hospital II. The patients from hospital I were divided randomly at a ratio of 8:2 into a training cohort and an internal validation cohort. The patients from hospital II were assigned to the external test cohort. The clinical characteristics of the three cohorts are summarized in Table 1. There were no significant differences among the three cohorts in any of the characteristics (all P > 0.05).

Table 1

Baseline data of the 3 cohorts
Characteristics	Training cohort (n = 71)	Internal validation cohort (n = 18)	External validation cohort (n = 41)	P value
Age (mean ± SD)	46.7 ± 14.7	42.9 ± 13.9	43.7 ± 10.3	0.381
Sex, n (%)				0.787
Male	3(4.2%)	1(5.6%)	3(7.3%)
Female	68(95.8%)	17(94.4%)	38(92.7%)
Histopathological result, n (%)				0.786
Without PTC	27(38.0%)	6(33.3%)	13(31.7%)
With PTC	44(62.0%)	12(66.7%)	28(68.3%)

Feature selection

A total of 851 radiomic features were extracted for each ROI. Among them, 750 features (87.9%) with high reproducibility were retained. After Pearson correlation analysis, 163 features were selected. Finally, 6 features with nonzero coefficients were retained by the LASSO algorithm for further ML model construction. The details of the features are shown in Fig. 3.

Model performance

Table 2 and Fig. 4 show the performance of the ML models. In the external validation cohort, the LR, NB, SVM, KNN and MLP models had AUCs of 0.736, 0.690, 0.751, 0.633 and 0.783, respectively. Among them, the MLP showed the best performance, with an AUC of 0.783 and a sensitivity and specificity of 64.29% and 92.31%, respectively, in the external validation cohort.

Table 2

Performance of the models in the identification of papillary thyroid carcinoma in Hashimoto's thyroiditis patients.
Cohort	Model	AUC (95%CI)	ACC (%)	SEN (%)	SPE (%)	PPV (%)	NPV (%)
Training cohort	LR	0.762 (0.647–0.876)	71.83	79.55	59.26	76.09	64.00
	NB	0.792 (0.684–0.901)	76.06	90.91	51.85	75.47	77.78
	SVM	0.866 (0.776–0.956)	81.69	86.36	74.07	84.44	76.92
	KNN	0.865 (0.785–0.945)	73.24	61.36	92.59	93.10	59.52
	MLP	0.796 (0.690–0.903)	71.83	70.45	74.07	81.58	60.61
Internal validation	LR	0.778 (0.458-1.000)	77.78	75.00	83.33	90.00	62.50
cohort	NB	0.819 (0.496-1.000)	83.33	83.33	83.33	90.91	71.43
	SVM	0.750 (0.428-1.000)	66.67	58.33	83.33	87.50	50.00
	KNN	0.625 (0.364–0.886)	44.44	16.67	1.00	1.00	37.50
	MLP	0.778 (0.457-1.000)	72.22	66.67	83.33	88.89	55.56
External validation	LR	0.736 (0.550–0.922)	75.61	75.00	76.92	87.50	58.82
cohort	NB	0.690 (0.497–0.882)	65.85	60.71	76.92	85.00	47.62
	SVM	0.751 (0.589–0.914)	68.29	64.29	76.92	85.71	50.00
	KNN	0.633 (0.439–0.827)	46.34	39.29	61.54	68.75	32.00
	MLP	0.783 (0.625–0.941)	73.17	64.29	92.31	94.74	54.55
*AUC, area under the receiver operating characteristic curve; ACC, accuracy; SEN, sensitivity; specificity. CI, confidence interval; PPV, positive predictive value; NPV, negative predictive value

The DCA of the 5 models in the external validation cohort demonstrated that the models had some benefit in most cases. The DCA of the MLP revealed that the model was clinically useful (Fig. 5).

Identifying PTCs in patients HT is crucial for therapeutic decision making to ultimately ensure timely and effective management. This study developed a radiomic model to predict the presence of PTC in patients with HT for early diagnosis. To our knowledge, this is the first study to predict the presence of PTC against the background of diffuse thyroid changes caused by HT using NECT radiomics. Six features were selected for ML model construction. The MLP model had the best performance, with an AUC of 0.783 and a sensitivity and specificity of 64.29% and 92.31%, respectively, in the external validation cohort.

Prior studies have supported the feasibility of radiomic approaches in PTC evaluation. Colakoglu et al. constructed a random forest model based on ultrasound texture features for distinguishing benign and malignant thyroid nodules with an AUC of 0.92¹³. Some researchers have attempted to predict occult lymph node metastasis in clinically negative lymph node PTCs by establishing a model based on CT radiomic features, suggesting that the radiomic signature can predict lymph node status by reflecting the microstructure of the primary lesion¹⁴. Wang et al. reported that multiparametric MRI-based radiomics can accurately distinguish aggressive from nonaggressive PTC¹⁵. These studies provide evidence the feasibility of radiomics in exploring the heterogeneity of thyroid lesions.

This study was not the first to analyze thyroid carcinoma in HT patients. Fang et al. proposed a radiomics model based on ultrasound to distinguish benign and malignant nodules in HT and achieved good results¹²; however, our study has several differences. First, we used NECT rather than ultrasound as the input. Both ultrasound and CT are commonly used imaging methods for evaluating the thyroid gland in clinical practice, but the quality of acquired ultrasound images is largely related to the doctor's personal experience, whereas CT is able to maximize the likelihood of acquiring consistent images through standardized scanning protocols¹⁶. Second, the previous study mainly extracted features from specific thyroid nodules. Considering that HT appears a diffuse lesion involving the entire thyroid gland, the present study extracted features of the entire thyroid gland and may be able to maximize the quantification of the entire thyroid gland in the context of HT.

Five ML models were constructed in this study. The MLP model had the best performance, with AUCs of 0.778 and 0.783 in the internal and external validation cohorts, respectively. The sensitivity, specificity, and accuracy of the MLP model in the external validation cohort were 64.3%, 92.3%, and 73.2%, respectively. While the performance of this model did not fully meet our expectations, its high sensitivity suggests that it may help us find patients with HT with suspected PTC. Furthermore, the increased use of chest CT, which is widely recommended for populations that require lung cancer screening, has resulted in increased detection of incidental thyroid disease^17–19. The models constructed based on the NECT radiomic features in this study also indicate the potential to detect incidental PTCs in patients with HT on chest CTs during lung screening.

Several limitations of our study should be acknowledged. First, the retrospective design and relatively small sample size may limit the generalizability of our findings. Second, significant selection bias exists because only surgically resected cases were included, resulting in a high proportion of positive PTCs. Finally, our study developed a radiomics model based only on CT images. Future studies should incorporate clinical factors to enhance the model's predictive value and comprehensiveness.

Radiomics models based on neck noncontrast enhancement CT could identify masked papillary thyroid carcinomas that are easily masked in diffuse Hashimoto's thyroiditis and are expected to facilitate the early diagnosis and intervention of thyroid carcinoma.

HT, Hashimoto's thyroiditis；

PTC, papillary thyroid carcinoma；

CT, computed tomography；

ML, machine learning；

NECT, nonenhanced CT；

LASSO, least absolute shrinkage and selection operator；

LR, logistic regression;

NB, naive Bayes;

SVM, support vector machine;

KNN, K nearest neighbor;

ANOVA, one-way analysis of variance;

ROC, receiver operating characteristic；

AUC, area under the curve；

DCA, Decision curve analysis.

Ethics approval and consent to participate

The Second Affiliated Hospital of Nanchang University Medical Research Ethics Committee approved the study and the waived the need for informed consent.

Consent for publication

Not applicable.

Availability of data and materials

The data that support the findings of this study are available from the corresponding author on reasonable request.

Competing interests

No competing financial interests exist.

Funding

This study has received funding by the National Natural Science Foundation of China (No. 91230027).

Contribution statement

(1) Concept and design by LG and YP. (2) Acquisition, imaging processing, statistical analysis and model construction by KH, ZG, WL and YP. (3) Drafting of the article by YP. (4) Critical revision of the article for important intellectual content by JP and L G.

(5) Funding obtained by LG.

Acknowledgments

We thank the Onekey AI platform for its services.

Ralli M, Angeletti D, Fiore M, et al. Hashimoto's thyroiditis: An update on pathogenic mechanisms, diagnostic protocols, therapeutic strategies, and potential malignant transformation. Autoimmun Rev 2020;19(10):102649, doi:10.1016/j.autrev.2020.102649
Xu J, Ding K, Mu L, et al. Hashimoto's Thyroiditis: A "Double-Edged Sword" in Thyroid Carcinoma. Front Endocrinol (Lausanne) 2022;13(801925, doi:10.3389/fendo.2022.801925
Xu S, Huang H, Qian J, et al. Prevalence of Hashimoto Thyroiditis in Adults With Papillary Thyroid Cancer and Its Association With Cancer Recurrence and Outcomes. JAMA Netw Open 2021;4(7):e2118526, doi:10.1001/jamanetworkopen.2021.18526
Uhliarova B, Hajtman A. Hashimoto's thyroiditis - an independent risk factor for papillary carcinoma. Braz J Otorhinolaryngol 2018;84(6):729-735, doi:10.1016/j.bjorl.2017.08.012
Lubin D, Baraban E, Lisby A, et al. Papillary Thyroid Carcinoma Emerging from Hashimoto Thyroiditis Demonstrates Increased PD-L1 Expression, Which Persists with Metastasis. Endocr Pathol 2018;29(4):317-323, doi:10.1007/s12022-018-9540-9
Ehlers M, Schott M. Hashimoto's thyroiditis and papillary thyroid cancer: are they immunologically linked? Trends Endocrinol Metab 2014;25(12):656-64, doi:10.1016/j.tem.2014.09.001
Vargas-Uricoechea H. Autoimmune Thyroid Disease and Differentiated Thyroid Carcinoma: A Review of the Mechanisms That Explain an Intriguing and Exciting Relationship. World J Oncol 2024;15(1):14-27, doi:10.14740/wjon1728
Zolotov S. Genetic Testing in Differentiated Thyroid Carcinoma: Indications and Clinical Implications. Rambam Maimonides Med J 2016;7(1), doi:10.5041/RMMJ.10236
Han ZJ, Shu YY, Lai XF, et al. Value of computed tomography in determining the nature of papillary thyroid microcarcinomas: evaluation of the computed tomographic characteristics. Clin Imaging 2013;37(4):664-8, doi:10.1016/j.clinimag.2012.12.005
Huang YQ, Liang CH, He L, et al. Development and Validation of a Radiomics Nomogram for Preoperative Prediction of Lymph Node Metastasis in Colorectal Cancer. J Clin Oncol 2016;34(18):2157-64, doi:10.1200/JCO.2015.65.9128
Zheng X, Yao Z, Huang Y, et al. Deep learning radiomics can predict axillary lymph node status in early-stage breast cancer. Nat Commun 2020;11(1):1236, doi:10.1038/s41467-020-15027-z
Fang M, Lei M, Chen X, et al. Radiomics-based ultrasound models for thyroid nodule differentiation in Hashimoto's thyroiditis. Front Endocrinol (Lausanne) 2023;14(1267886, doi:10.3389/fendo.2023.1267886
Colakoglu B, Alis D, Yergin M. Diagnostic Value of Machine Learning-Based Quantitative Texture Analysis in Differentiating Benign and Malignant Thyroid Nodules. J Oncol 2019;2019(6328329, doi:10.1155/2019/6328329
Peng Y, Zhang ZT, Wang TT, et al. Prediction of Central Lymph Node Metastasis in cN0 Papillary Thyroid Carcinoma by CT Radiomics. Acad Radiol 2023;30(7):1400-1407, doi:10.1016/j.acra.2022.09.002
Wang H, Song B, Ye N, et al. Machine learning-based multiparametric MRI radiomics for predicting the aggressiveness of papillary thyroid carcinoma. Eur J Radiol 2020;122(108755, doi:10.1016/j.ejrad.2019.108755
Wu X, Li J, Mou Y, et al. Radiomics Nomogram for Identifying Sub-1 cm Benign and Malignant Thyroid Lesions. Front Oncol 2021;11(580886, doi:10.3389/fonc.2021.580886
Seo H, Jin KN, Park JS, et al. Risk of thyroid cancer in a lung cancer screening population of the National Lung Screening Trial according to the presence of incidental thyroid nodules detected on low-dose chest CT. Ultrasonography 2023;42(2):275-285, doi:10.14366/usg.22111
van Gerwen M, Ma T, Yip R, et al. Thyroid abnormalities identified on CT screening for lung cancer. Clin Imaging 2024;110(110162, doi:10.1016/j.clinimag.2024.110162
Mazzone PJ, Silvestri GA, Souter LH, et al. Screening for Lung Cancer: CHEST Guideline and Expert Panel Report. Chest 2021;160(5):e427-e494, doi:10.1016/j.chest.2021.06.063

No competing interests reported.

Download PDF

Editorial decision: Revision requested
14 Aug, 2024
Editor assigned by journal
06 Aug, 2024
Submission checks completed at journal
06 Aug, 2024
First submitted to journal
05 Aug, 2024

You are reading this latest preprint version

Unveiling Hidden Carcinomas: nonenhanced CT-Based Radiomics Model Enhances PTC Detection in Hashimoto's Thyroiditis Running title: CT radiomics identifies carcinoma in HT

Status:

Version 1

Abstract

Figures

Introduction

MATERIALS AND METHODS

Patients

CT scanning

Segmentation of the region of interest and extraction of radiomic features

Feature selection and model establishment

Statistical analysis and model performance evaluation

RESULT

Clinical characteristics

Feature selection

Model performance

DISCUSSION

Conclusion

Abbreviations

Declarations

References

Additional Declarations

Status:

Version 1