In recent decades, there has been a significant global rise in the incidence of thyroid cancer. The majority (over 80%) of cases belong to the papillary thyroid carcinoma (PTC) subtype (1). The 2015 American Thyroid Association Guidelines introduced active surveillance as a safe alternative to surgical intervention for low-risk PTCs. However, for high-risk PTCs, aggressive surgical treatment remains the only reliable and safe option (2). The BRAF^V600E gene mutation is closely associated with PTC and plays a crucial role in its pathogenesis. Previous study has demonstrated that PTC cases with the BRAF^V600E mutation exhibited a more aggressive behavior than those without such mutation (3). Therefore, accurately assess the BRAF^V600E mutation status is critically important for definitive treatment decisions.

Currently, the BRAF^V600E mutation tests are typically performed on preoperative invasive fine-needle aspiration biopsy samples or surgically resected specimens. However, the use of ultrasound-guided fine needle puncture is invasive and can lead to potential adverse effects, such as pain, bleeding, and acute thyroid swelling (4–6). Texture analysis, a well-established technique, is applied when performing quantitative analysis of tumor heterogeneity by analyzing the distribution and the relationship of pixels or voxel gray levels in the target region (7). The ease of obtaining texture information from routinely acquired images without additional imaging procedures is required, and accumulating data showing the correlation between heterogeneity and adverse tumor biology is a major advantage of the technique, which has been utilized to identify tumor differentiation degrees, assess tumor characteristics, and evaluate treatment efficacy (8, 9). While there have been limited studies on the use of texture analysis for predicting the molecular state of thyroid cancer, most of these studies have focused on ultrasound images (10, 11). Although ultrasound images exhibit high sensitivity and specificity in diagnosing thyroid cancers, there were some limitations such as operator and machine dependence, limited image data, and a lack of representative features. Conversely, CT imaging provides more intuitive and accurate information directly from CT scans. Additionally, CT imaging is particularly effective in visualizing calcification, especially large and marginal calcifications (12). However, when performing texture analysis, regions of interest (ROI) should be delineated while excluding calcifications, despite their presence in certain thyroid tumors and their significance in distinguishing between benign and malignant tumors (13). Further investigation is thus needed to determine whether the presence of calcifications influences the results of texture analysis.

The primary objective of this study is to investigate the potential of texture analysis base on the preoperative CT imaging, in predicting the presence of BRAF^V600E mutation in calcified PTC. Additionally, our study aims to compare the diagnostic performance of various texture feature extraction methods.

Our study revealed that the utilization of MaZda software for texture analysis of preoperative CT images of PTCs enables the prediction of the BRAF^V600E mutation. Notably, when considering either the parenchymal or calcified areas of the tumor, the NDA method exhibited excellent diagnostic performance in conjunction with any of the feature selection algorithms. Specifically, the combination of the POE+ACC algorithm with the NDA method, as well as the combination of the MI algorithm with the NDA method, demonstrated the highest diagnostic performance in predicting the BRAF^V600E mutation in PTC. Furthermore, we meticulously delineated ROIs in both the parenchymal and the calcified areas of the tumor. Remarkably, our findings indicated that there was no significant difference in diagnostic performance between these two distinct areas.

PTC patients with BRAF^V600E positivity exhibited more aggressive clinical behavior, as determined by the constitutive abnormal activation of the mitogen-activated protein kinase (MAPK) signaling pathway driven by the BRAF^V600E mutation, along with its downstream cascade of complex molecular and cellular biological effects (16, 17). In addition, PTC displays significant gender disparity. Epidemiological studies have demonstrated that its occurrence rate in females is approximately 3–4 times higher than in males, which aligns with our study’s female-to-male ratio of approximately 3:1 (18). This phenomenon may be attributed to sex hormone (particularly estrogen)-mediated activation of oncogenic pathways, X chromosome-driven establishment of a gender-specific tumor microenvironment, and differences in healthcare-seeking behaviors between sexes (18–20).

In this study, the freely available software MaZda was use for feature extraction and dimensionality reduction. MaZda provides several feature extraction methods, including Fisher, POE+ACC, and MI, along with dimensionality reduction techniques such as PCA, LDA, and NDA (21, 22). PCA evaluates the covariance matrix of data, while LDA calculates the scattering matrix. Both methods use linear transformation for data linear classification. However, although PCA can effectively represent the dataset, its performance in data classification is relatively subpar (23). Our study supports these findings, as we found that the PCA method yielded unreliable results. Similarly, while LDA demonstrated accuracy in certain data classifications, it imposed high requirements on data characteristics, resulting in overall suboptimal diagnostic performance. Conversely, most datasets in our study consisted of non-linear data, which can be better classified using non-linear classifiers such as NDA instead of linear classifiers. NDA employs neural networks to perform non-linear transformations on data, projecting them into lower-dimensional feature spaces for linear classification (24, 25). Given the tumor heterogeneity, partial tumor cells serve as the basis for BRAF^V600E mutation. Additionally, BRAF^V600E mutation is associated with various factors, including tumor size, multifocality, tumor stage, aggressiveness, and the presence of Hashimoto’s thyroiditis (26, 27). Consequently, we can conclude that NAD is more suitable than the other two methods for predicting BRAF^V600E mutation, as it achieves higher accuracy by effectively modeling the non-linear characteristics of the dataset collected from BRAF^V600E mutations. Our study demonstrated that NDA exhibited excellent diagnostic performance, with an error rate of less than 10% in both the parenchymal and calcified areas, irrespective of the feature selection algorithm utilized.

Previous research has demonstrated that the MaZda software holds promise in distinguishing between benign and malignant tumors. For instance, Ardakani et al. (28) revealed that a combination of POE+ACC with NDA yielded excellent diagnostic performance in differentiating thyroid nodules using ultrasound images, achieving an accuracy of 97.14%, which closely aligns with our findings. However, their study only employed the Fisher and POE +ACC algorithms. On the other hand, our study revealed that the combination of MI with NDA also exhibited excellent diagnostic performance, with an AUC of 0.969. In contrast, Kwon et al. (10) observed that various dimensionality reduction and separator techniques had minimal influence on the results, while the choice of feature selection algorithm significantly impacted performance. Consequently, they concluded that radiomics studies using thyroid sonography have limited predictive ability for the BRAF mutation status of PTC. Yoon et al. (11) reported similar findings. Our study demonstrated that the NDA method achieved high accuracy, with an AUC exceeding 0.9. We hypothesize that the utilization of different radiomics software may have contributed to the discrepancies, highlighting one of the major challenges in current radiomics research. Furthermore, due to the diversity of texture feature analysis methods, the results can vary depending on the chosen methods, even within the same software. Therefore, radiomics research must establish unified and standardized protocols.

The majority of previous radiomics studies have emphasized the avoidance of calcification when delineating ROIs. However, it is worth noting that calcification not only is prevalent in thyroid tumors but also serves as a significant differential diagnostic indicator. While ultrasound plays a crucial role in thyroid tumor diagnosis, the elimination of calcification when delineating ROI, particularly in cases of microcalcification, poses challenges. Currently, whether calcification in thyroid tumors should be excluded during ROI delineation remains uncertain. For example, some previous studies have included calcification when drawing the ROIs and have discovered that ultrasound and MRI radiomics can be used to evaluate thyroid nodules (28–30). Conversely, other studies have suggested that calcification should be excluded (31, 32). In our study, we delineated the ROI separately in the parenchymal and calcified areas of the tumor. We confirmed that drawing the ROI in the calcified area can be also used to predict the BRAF^V600E mutation. There was no significant difference in diagnostic performance between these two areas. Two potential explanations may account for this finding. Firstly, the neovascularization and fibrous tissue in thyroid tumors are prone to cause calcium salt deposition due to the rapid proliferation of cancer cells. Additionally, the tumor itself secretes glycoprotein and mucopolysaccharide, which may contribute to nodule calcification. Secondly, microcalcification can be observed in 50% PTC cases and has been confirmed as an important factor associated with the BRAF^V600E mutation in PTC (26, 33).

This study had several limitations. Firstly, the data were processed using the default format of the Mazda software. Whether this affects the results requires further investigation. Secondly, we only focused on nodules larger than 5 mm and excluded those with thyroiditis or other inflammatory lesions. Furthermore, the mutation analysis was performed in a specific subgroup of PTCs. Therefore, our study did not encompass the mutation features of other types of thyroid cancer. Thirdly, texture analysis was performed on venous-phase images, as tumor tissues exhibit clearer boundaries and more stable texture features during this phase. Considering the potential influence of varying delay times on texture features and their diagnostic performance, a standardized delay of 45–50 seconds was employed for venous-phase imaging in this study. This aspect warrants further investigation. Finally, the retrospective design introduced variability in CT scanning parameters, which may have affected feature extraction and introduced measurement bias. Incomplete clinical data, such as thyroid function laboratory indicators, also somewhat limited the generalizability of the results. Future investigations will address these factors through prospective, multicenter studies to validate the present findings.

In conclusion, our findings demonstrated that texture analysis based on preoperative CT images of PTCs using MaZda software can predict the presence of the BRAF^V600E mutation. The combination of POE+ACC with the NDA method or MI with the NDA method exhibited the highest diagnostic performance in predicting the BRAF^V600E mutation in PTC. The calcified area of the tumor can be utilized for predicting the BRAF^V600E mutation, and there was no significant difference in diagnostic performance between the parenchymal and calcified areas. Nevertheless, our results should be further validated in a larger sample size to enhance their potential clinical utility.