CT-based texture analysis predicts BRAFV600E mutation in calcified papillary thyroid carcinoma

TL;DR

This study shows that CT-based texture analysis can accurately predict BRAFV600E mutations in calcified papillary thyroid carcinoma.

Contribution

The novel use of texture analysis in calcified tumor areas to predict BRAFV600E mutations in papillary thyroid carcinoma.

Findings

Nonlinear discriminant analysis (NDA) achieved high diagnostic performance with low error rates.

Texture analysis of calcified areas was effective in predicting BRAFV600E mutations.

POE+ACC+NDA and MI+NDA methods showed the best performance with AUC values above 0.96.

Abstract

The BRAF gene plays an essential role in papillary thyroid carcinoma (PTC).

To investigate the potential of CT-based texture analysis in predicting BRAFV600E mutation in calcified PTC.

475 cases of calcified PTC from two centers, who underwent CT scans, surgery, and BRAFV600E mutation testing, were included. Data from the first center were randomly divided into training and testing sets, whereas data from the second center constituted an external validation set. Using MaZda software, 256 texture features were extracted from both the parenchymal and calcified areas. The top ten texture feature parameters were selected by Fisher, minimization of both classification error probability and average correlation coefficients (POE+ACC), and mutual information measure (MI) feature selection algorithms. Data analysis and classification were performed using principal component analysis (PCA), linear discriminant analysis (LDA), and nonlinear discriminant analysis (NDA). Receiver operating characteristic curves were used to evaluate the diagnostic performance.

The NDA method demonstrated excellent diagnostic performance compared to the LDA and PCA methods, with error rates of less than 10%, less than 25%, and greater than 30%, respectively in the training and validation sets. For parenchymal and calcified areas of PTC, the POE+ACC+NDA and MI+NDA methods exhibited the lowest error rates, with an area under the curve (AUC) of 0.969 in the training set and 0.964 in the internal validation set. Conversely, the Fisher+PCA and MI+PCA methods had the highest error rates, with AUC values of 0.413 and 0.525 in the training set, and 0.433 and 0.560 in the internal validation set, respectively.

The POE+ACC+NDA or MI+NDA method provided high diagnostic performance for predicting BRAFV600E mutation in PTC. Texture analysis of tumor calcified area can also be used to predict BRAFV600E mutation.

Figures