Introduction
Recent data suggest that the increasing incidence of thyroid cancer may be starting to slow [1]; however, the incidence of thyroid cancer in several developed countries has markedly increased [2, 3]. In 2019, new cases of thyroid cancer account for 4% of all new cancers diagnosed in women [4], ranking sixth in the world. Although thyroid cancer has a good prognosis, the incidence of cervical lymph node metastasis (CLNM) can reach 30% to 80% [5, 6]. Studies have shown that lymph node metastasis is related to thyroid cancer recurrence [7, 8]; thus, early identification or prediction of CLNM in thyroid cancer is essential.
Artificial intelligence has been widely used in various fields and has quickly penetrated many aspects of modern life. In recent years, artificial intelligence has made rapid progress in the medical field [9], and there are a number of research studies have examined deep learning on massive data [10]. Artificial intelligence may provide an early diagnostic strategy for lymph node metastasis in patients with breast cancer with clinically negative lymph nodes [11]. Other aspects of medicine require more research based on artificial intelligence, for example, papillary thyroid carcinoma (PTC). Artificial intelligence has been used to optimize the Thyroid Imaging Reporting and Data System (TI-RADS) to validate the American College of Radiology TI-RADS and slightly improve specificity and maintain sensitivity [12]. A large number of data have verified the risk factors of CLNM in PTC, and the overall factors are consistent, including age, sex, maximum diameter, aspect ratio, and calcification [2, 13–15]. Although the use of artificial neural networks (ANNs) in predicting central lymph node metastases in papillary thyroid microcarcinoma had been researched, the prediction of lymph node metastases in PTC by a multilayer perceptron (MLP) has not been studied and needs further exploration [16]. At the same time, the MLP has many forms of charts that can show the important factors in predicting the outcome intuitively. This study aimed to explore the risk factors of lymph node metastasis according to preoperative thyroid ultrasound characteristics and clinically relevant patient information and to predict the potential of CLNM in PTC by using neural network MLP, thus guiding individualized clinical treatment.
Materials and methods
Patients
Information on patients with thyroid nodules who were treated at Shengjing Hospital of China Medical University and had pathology results from January 2018 to October 2019 was collected. A total of 1139 nodules from 896 patients were collected. Patients’ clinical information was collected through the HIS system, including information such as sex, age, family history, history of radiotherapy, and thyroid function tests. All patients underwent cervical ultrasound within 7 days prior to surgery, and the metastasis of both central and lateral cervical lymph nodes was recorded. Patients who had thyroidectomy or missing statistical data, those whose nodules were benign or who had other pathological types of thyroid cancer, and patients younger than 18 years were excluded. In total, 478 PTC nodules from 383 patients were collected in our statistical study, forming a neural network model ( Figure 1 ). After forming the model, the data for 60 new patients with PTC undergoing surgery from January to May 2020 in Shengjing Hospital of China Medical University were evaluated with the previously established model.
Ultrasonic classification criteria
Ultrasonic testing of patients with PTC was performed before the operation by two ultrasound doctors with >5 years of experience. The ultrasound characteristics of thyroid nodules were recorded and analyzed in the following aspects: nodular echo (hypoechoic and equal/hyperechoic), size (≤1 and >1 cm), location (upper, middle, and lower poles), number (single and multiple), depth (close or not close to the envelope), tissue background (normal and Hashimoto’s thyroiditis), aspect ratio (>1 and ≤1), boundary (clear and unclear), shape (regular and irregular), calcification (calcification and non-calcification), blood distribution (detectable and undetectable), and posterior attenuation (with and without). According to the classification of the maximum diameter of nodules (≤1 or >1 cm), PTC with a maximum dimension <1 cm was differentiated as microcarcinoma [17, 18].
Pathological diagnosis
The lymph node status of each PTC in this study was obtained from the postsurgical pathological diagnosis report in our hospital. PTC with CLNM was defined as CLNM, and PTC without metastasis was defined as non-CLNM.
Statistical analysis
SPSS version 22.0 (IBM Corp., Armonk, NY, USA) was used in the statistical analysis. Logistic regression analysis was used to explore the different factors associated with CLNM, which was the criterion to predict the CLNM of PTC. A P value <0.05 was considered statistically significant.
The MLP, which is called an ANN, was used to train and test the veracity to predict the CLNM as a new method. The samples were split in the neural network by the default settings of SPSS. The architecture of the MLP model includes hidden and output layers; in the module of the hidden layers, the model had a hidden layer, and the activation function was chosen as the hyberbolic tangent and was automatically computed, whereas in the output layer, the activation function was identity. We formed a training set and a verification set according to the known data in a 7:3 ratio. The closer the areas under the receiver operating characteristic (ROC) curve are to 1, the higher the accuracy of the assessment. MedCalc software (MedCalc, Ostend, Belgium) was used to compare the metastasis status predicted by this neural network model with logistic regression. According to the MLP neural network model with .xml format constructed by the known data, the risk of metastasis of the 60 new patients with PTC was evaluated on the basis of the previously established model. Sex, age, maximum diameter, location, number, depth, tissue background, aspect ratio, boundary, shape, calcification, blood distribution, posterior attenuation, echo, family history, and thyroid function were used as factors evaluating lymph node metastasis. The metastasis status of these patients was then compared with the pathological results.
Results
Patient information and ultrasonic characteristics of thyroid nodules
The clinical information on 383 patients with 478 nodules of PTC were collected from our department, and the ultrasound characteristics of all the thyroid nodules were received before surgery. There were 309 (80.7%) female and 74 (19.3%) male patients. The age of the patients ranged from 22 to 80 years, and the average age was 45.8 years. A total of 214 (44.8%) nodules were CLNM, and 264 (55.2%) were non-CLNM ( Table 1 ).
Sex | |
Female | 385 (80.5) |
Male | 93 (19.5) |
Lymph node | |
Non-CLNM | 264 (55.2) |
CLNM | 214 (44.8) |
Family history | |
No | 452 (94.6) |
Yes | 26 (5.4) |
Thyroid function | |
Normal | 259 (54.2) |
Abnormal | 219 (45.8) |
Values are presented as frequency (percentage).
CLNM, cervical lymph node metastasis.
After integration of multiple factors, including the basic information on patients and the ultrasonic characteristics of thyroid nodules, the factors related to PTC were determined by multiple regression. Sex, age, thyroid function of the patients, maximum diameter, aspect ratio, and calcification of the thyroid nodules were the risk factors of CLNM ( Table 2 ). Male patients with PTC were 1.87 times more likely to have CLNM than women (P<0.05). Patients younger than 40 years were 3.21 times more likely to have CLNM than patients 40–60 years, who, in turn, were 2.72 times more likely to have CLNM than patients older than 60 years. Thyroid nodules with maximum diameter >1 cm had 1.86 times higher risk of CLNM than those with a diameter ≤1 cm. Patients with an aspect ratio ≤1 cm (odds ratio [OR]=1.58, P<0.05), calcification (OR=2.29, P<0.05), and abnormal thyroid function (OR=1.66, P<0.05) had a higher risk of CLNM than non-CLNM.
Metastasisa | B | Standard Error | Wald | Sig. | Exp(B) | 95% Confidence Interval for exp(B) | |
---|---|---|---|---|---|---|---|
Lower Bound | Upper Bound | ||||||
Intercept | 0.180 | 0.866 | 0.043 | 0.836 | |||
Sex | –0.626 | 0.264 | 5.629 | 0.018* | 0.535 | 0.319 | 0.897 |
Age | |||||||
<40 years | 1.166 | 0.278 | 17.641 | <0.0001* | 3.209 | 1.862 | 5.529 |
40–60 years | 1.000 | 0.274 | 13.378 | <0.0001* | 2.720 | 1.591 | 4.649 |
Location | |||||||
Upper pole | –0.328 | 0.271 | 1.467 | 0.226 | 0.720 | 0.423 | 1.225 |
Middle pole | –0.091 | 0.235 | 0.152 | 0.697 | 0.913 | 0.576 | 1.446 |
Number pole | –0.338 | 0.316 | 1.144 | 0.285 | 0.713 | 0.384 | 1.325 |
Depth (close to the envelope) | 0.115 | 0.228 | 0.254 | 0.614 | 1.122 | 0.718 | 1.753 |
Tissue background (normal) | 0.011 | 0.347 | 0.001 | 0.975 | 1.011 | 0.512 | 1.998 |
Maximum diameter | –0.620 | 0.231 | 7.171 | 0.007* | 0.538 | 0.342 | 0.847 |
Aspect ratio | 0.457 | 0.227 | 4.044 | 0.044* | 1.579 | 1.012 | 2.464 |
Boundary | 0.127 | 0.289 | 0.193 | 0.660 | 1.135 | 0.644 | 2.000 |
Shape | –0.025 | 0.231 | 0.012 | 0.913 | 0.975 | 0.620 | 1.533 |
Calcification | –0.830 | 0.230 | 13.021 | <0.0001* | 0.436 | 0.278 | 0.684 |
Blood distribution | 0.258 | 0.222 | 1.349 | 0.245 | 1.294 | 0.838 | 2.000 |
Posterior attenuation | 0.138 | 0.232 | 0.354 | 0.552 | 1.148 | 0.729 | 1.809 |
Echo | 0.266 | 0.497 | 0.285 | 0.593 | 1.304 | 0.492 | 3.456 |
Family history | –0.312 | 0.427 | 0.532 | 0.466 | 0.732 | 0.317 | 1.691 |
Thyroid function tests | –0.510 | 0.214 | 5.652 | 0.017* | 0.601 | 0.395 | 0.914 |
*P<0.05.
aOccurred neck lymph node metastasis.
CLNM, cervical lymph node metastasis.
Factors related to lymph node metastasis by the MLP
Calcification, sex, age, and maximum diameter could be used to predict lymph node metastasis with MLP ( Figures 2 and 3 ). During the training stage, the percent correction of predicting non-CLNM was 75.4% compared with the pathology based on ultrasonic characteristics and clinical information. During the verification stage, the consistency predicting non-CLNM was 76.5% according to the training results ( Table 3 ). The area under the ROC curve for this method of predicting metastasis was 0.715 (P<0.001) ( Figure 4 ), thus indicating that it can meaningfully distinguish between lymph node metastasis and non-metastasis.
Comparison of the area under ROC curves on multiple regression and the MLP
The area under the ROC curve of multiple regression was 0.711, whereas that of MLP was 0.715 with MedCalc. The P value was 0.69 (Table 4), thus indicating that no significant differences were found between the MLP and multiple regression in predicting CLNM.
Prediction of new patient metastasis with the known model
For 60 patients with PTC different from the previous patients, 36 cases were predicted to be consistent with the pathological results of the thyroid nodules, and 24 cases were different, according to the basic information of the patients and the ultrasound characteristics of the thyroid nodules. The average confidence of the model used in these new patients for predicting metastasis with PTC was 68.9%.
Discussion
Thyroid cancer has a higher proportion of CLNM [5]. However, the probability of preoperative routine ultrasound or computed tomography examination to detect metastatic lymph nodes is not sufficiently high [19, 20]. Therefore, we wanted to predict CLNM according to clinical information and the preoperative ultrasound characteristics of the thyroid nodules in order to provide a reference for surgery.
According to our data, patient sex, age, and thyroid function were significantly associated with CLNM by traditional multifactor analysis. We found that the patients with aspect ratio ≤1 cm (OR=1.59, P<0.05) more frequently had CLNM than non-CLNM. In a study [21] summarizing 1179 nodules, when the aspect ratio was ≤1, the nodules were mostly benign. In other words, an aspect ratio >1 may be a risk factor for malignant nodules, but among all malignant nodules, the number of nodules between the two groups with aspect ratios >1 and ≤1 was similar. Studies [22, 23] have also observed no significant relationship between aspect ratio and CLNM of thyroid cancer. A study [24] has shown that an aspect ratio >1 is a risk factor for central lymph node metastasis with Hashimoto’s thyroiditis. The conclusions are slightly different among different studies. This conclusion is related to the different backgrounds of thyroid nodules, to the limitation of thyroid papillary microcarcinoma, and to the limitation of central lymph nodes; thus, the conclusions of various documents differ. Thus, the influencing factors of PTC lymph node metastasis are not very powerful when the aspect ratio is >1, and thus require further study.
When the artificial intelligence MLP method was used to predict the veracity of CLNM, the clinical information and nodular ultrasound features of 478 PTC nodules were trained and tested. Seventy percent of the data were used for training to predict CLNM, and the remaining 30% were used to verify this prediction method. Cross contrasts the correct rate of the distinguished samples, showing that the probability of predicting non-CLNM was higher than that of CLNM, and the model’s ability to identify CLNM risk was lower than its ability to identify non-CLNM. Factors that affected the results by >50% were used as criteria for predicting lymph node metastasis, including calcification, age, sex, and maximum diameter.
Based on the above results, younger patients have a higher risk of metastasis. This may be related to body functions and metabolic conditions. Aging is the main risk factor for cancer progression, due to accumulation of mutations and a decline of homeostasis occurring during aging [25]. However, aging may not be a high-risk factor for metastasis. Paik et al. [26] believed that older patients have lower risk of recurrence of breast cancer, and the reason is not related to age itself but to the higher amount of estrogen-receptor protein in older patients’ breast tumors. A study has found a more progressive and higher recurrence rate of clinical PTC among young patients than older ones [27]. Therefore, a high probability of metastasis or recurrence of PTC at a young age may be related to hormone changes.
The maximum diameter and calcification factors of thyroid nodules were associated with CLNM in thyroid cancer in this study. The size of the largest diameter of thyroid nodules is related to CLNM. A study has shown that nodules >6 mm are a risk factor for metastasis in papillary thyroid microcarcinoma [28]. On the one hand, if the nodule is not found in the early stage and the treatment is not timely, the nodule may grow and have a high possibility of metastasis. On the other hand, some patients may choose regular follow-up instead of undergoing surgery. Suggestions should be given to patients and clinics to avoid thyroid lymph node metastasis when the nodules are >1 cm. Calcification is used as an American Thyroid Association guideline when evaluating thyroid nodules. Thyroid nodules with calcification have a higher risk of metastasis. It is believed that calcification is an independent risk factor for a large number of CLNM (number of metastatic lymph nodes >5) [29], and microcalcification is an indicator of rapid proliferation of cancer cells and may be a potential promoter of CLNM in PTC [30, 31].
SPSS adopts 0.5 as the probability boundary between right and wrong by default. In the independent variable importance map, the importance value was sorted in descending order. The proportion and importance of calcification, age, sex, and maximum diameter were seen when identifying CLNM. The area under the ROC curve was 0.715, thus indicating a predictive effect. Compared with binary logistic regression, the MLP method is clearer and more intuitive. The areas under the ROC curve between these two methods for predicting CLNM of thyroid cancer were 0.711 and 0.715, respectively. Both logistic regression and the MLP meet the needs for CLNM prediction. The areas under the curve were similar, but the MLP’s synaptic weights chart, lift chart, predicted versus observed chart, and other charts can more intuitively show the most important factors in predicting the outcome. The ANN method is not only more intuitive, but also more convenient and meaningful. By including the relevant ultrasound information of the patient’s nodule and clinical information, information can be obtained on whether the lymph node may metastasize. Currently, the model is not stable enough, and in further research, the sample size could be increased to improve its prediction veracity and stability.
According to the MLP neural network model, the risk of metastasis of 60 new patients with PTC was predicted, and the average confidence was 68.9%. To a certain extent, this model can be used to predict cervical lymph node metastases from PTC, but the accuracy of the prediction needs to be improved. In future studies, the data volume of the training model should be increased to make the predicting model more stable, and more new data should be collected for prospective research.