Recent investigations have shown that patient prognosis is comparatively worse in advanced (vs. limited) thyroid cancers [8]. Herein, we examined demographic and sonographic parameters of patients in T4 or M1 disease stages. We also evaluated patients with limited thyroid carcinomas (stage T3 or less, no distant metastases) for purposes of comparison. To avoid sampling bias, candidates with incidental papillary microcarcinomas were deemed ineligible.
In our patient population, those with advanced disease were on average older compared to others with less prolific cancers. Male patients also accounted for a higher proportion of subjects with advanced disease. Hwang et al. have likewise identified male sex as an independent risk factor for thyroid malignancy [20]. In addition, in clinicopathologic comparisons of various thyroid carcinomas, increasing median ages among patients with anaplastic, poorly differentiated, and differentiated carcinomas have been recorded [5, 21].
Regarding sonographic parameters, we found that tumor volumes in ADV group members significantly surpassed those of the limited disease groups. Overall, tumor size determined by ultrasound correlated well with measurements obtained during pathologic examination, although multinominal logistic regression analysis revealed a more than three-fold rise in the incidence of advanced (vs. limited) disease for tumors with irregular shapes and contours.
The impact of tumor size on risk of T4 disease stage or distant metastases has already been explored in an earlier study [13]. These authors found that in differentiated thyroid carcinomas, the risk of local invasion (T4) or distant spread (M1) increases gradually along with tumor size. Such increases appeared linear for PTCs (without threshold effect) and non-linear for FTCs beyond 4 cm in diameter. In terms of distant metastases, no size thresholds were evident for PTCs or FTCs, although the probability of distant metastases increased progressively with size in undifferentiated thyroid cancers.
The mean tumor diameter we determined for all types of ADVs was ~ 5.4 cm. Subgroup analysis further revealed mean tumor diameters of 1.7 cm and 3.1 cm for PTCs and FTCs, respectively. These findings imply that in the context of advanced thyroid cancers, no threshold values are definable for culpable primary tumors.
In our patients, irregularly shaped tumors were statistically more frequent in those with advanced (vs. limited) disease. A large-scale meta-analysis has also shown that irregular margins (among other features) are highly predictive of malignancy [14]. Unfortunately, the ultrasound features of advanced or non-advanced tumors were not addressed, particularly ramifications of round, oval, or irregular sonographic tumor shapes.
Hahn et al. have reported that shapes and margins of thyroid tumors on ultrasound may reflect levels of biologic aggression [22]. For instance, oval-to-round appearances and well-defined margins were detected more often in poorly differentiated carcinomas than in anaplastic tumors. These revelations perhaps support the significant disparities in irregular tumor shapes and margins exhibited by advanced and non-advanced tumors in the course of our multinomial logistic regression analysis.
Our investigation was not designed as an observer study. Therefore, we are unable to provide data on the intra- and interobserver variances of the ultrasound findings. Ultrasound examinations were carried out, and the results were classified by examiners with high experience in this field. We assume that this approach was feasible to keep the variances low.
The structure of neural networks, as well as training and validation processes, has been extensively described by Lee et al. [17]. They contend that this technology may help integrate the diagnostic intricacies of complex pathologies. Reliance on neural networks for quantitative data processing may indeed provide greater diagnostic accuracy in patients with suspected advanced cancerous lesions.
We used a two-step approach to evaluate our network. Training and validation were done alternately to ensure sufficient generalizability during training, testing predictive power on a hold-out dataset. Training was halted when the loss of function converged. Our proof-of-concept network led to correct classification in most patients (84%) with ADVs. Jeong et al. have evaluated a commercially available CAD system for ultrasonographic recognition of thyroid cancers [16], reaping a positive predictive value of 81.3%. However, these commercially available artificial intelligence systems were devoid of clinical input, restricted to ultrasound parameters only [23].
As a retrospective study, Li et al. recently examined the diagnostic performance of a deep convolutional network model to differentiate malignant and benign thyroid nodules based on ultrasound imaging data [24]. The observed accuracy of this model in correctly classifying respective lesions was also quite high (> 85%). Our smaller sampling achieved similar accuracy (84.4%) in discriminating advanced from limited thyroid cancers, thus indicating the high potential of adjunctive neural network learning methods in imaging analysis. Unlike the model of Li et al., our approach allows the implementation of a freely accessible online data input tool. Because voluminous data is not essential, our application is a practical one.
Besides the neural network that was used in our study to classify thyroid nodules, several other approaches using computer-aided diagnosis systems (CAD) have been evaluated. Wei et al. [25] compared the diagnostic value of S-Detect, a CAD system used to differentiate benign and malignant thyroid nodules by radiologists with different levels of experience. They reported that S-Detect had an accuracy, sensitivity, specificity, positive predictive value, and negative predictive value of 77.0, 91.3, 65.2, 68.3, and 90.1%, respectively. We could demonstrate a higher efficiency of our neural network approach. Furthermore, in contrast to our study, in Wei et al.’s study not all of the histopathological results were obtained by surgical resection; this might have further restricted the study results. Additionally, Kim et al. reported that S-Detect has a limitation in the evaluation of nodule calcifications, restricting its use in the evaluation of calcified thyroid nodules [26]. Xia et al. evaluated the use of S-Detect in 171 patients with 180 thyroid lesions [27]. They found that the CAD system presented a higher sensitivity but lower specificity than an experienced radiologist (90.5% vs. 81.1 and 41.2% vs. 83.5%). The radiologist also had a higher accuracy compared to the CAD system (82.2% vs. 67.2%) for diagnosing malignant thyroid nodules. The authors concluded that S-Detect had a lower specificity and accuracy than the experienced radiologist in identifying papillary thyroid carcinomas and also maintained a relatively lower performance than the experienced radiologist in identifying follicular thyroid carcinomas. Unlike S-Detect, our presented neural network approach allows the implementation of a freely accessible online data input tool that enables simple non-commercial use in the future.
There are several limitations to our study, the first being its retrospective design. We included only those thyroid tumors from our database that were identifiably encoded. Another issue is that only patients surgically treated at our facility with available pathologic reports were considered. Various protocols used were also clinically based and non-standardized, and the small number of patients involved who were not perfectly matched may have introduced significant outcome bias. Our investigation was performed as a single center retrospective study. One should be aware that this design might have reduced the statistical power of our results. A bias regarding the parameter tumor size as input function of the neural network cannot be excluded in our data. However, this parameter alone is not decisive as to whether the neural network classifies a tumor as advanced or not. In the advanced tumor group, 21 of 32 patients presented with distant metastases and were therefore included in the advanced group. Finally, we used a concise rather than comprehensive neural network model for analysis, requiring some simplification of output functions.