A Model Using Texture Features to Differentiate the Nature of Thyroid Nodules on Sonography.

OBJECTIVES

To evaluate the use of texture-based gray-level co-occurrence matrix (GLCM) features extracted from thyroid sonograms in building prediction models to determine the nature of thyroid nodules.

METHODS

A GLCM was used to extract the texture features of 155 sonograms of thyroid nodules (76 benign and 79 malignant). The GLCM features included energy, contrast, correlation, sum of squares, inverse difference moment, sum average, sum variance, sum entropy, entropy, difference variance, difference entropy, information measures of correlation, and maximal correlation coefficient. The texture features extracted by the GLCM were used to build 6 different statistical models, including support vector machine, random tree, random forest, boost, logistic, and artificial neural network models. The models' performances were evaluated by 10-fold cross-validation combining a receiver operating characteristic curve, indices of accuracy, true-positive rate, false-positive rate, sensitivity, specificity, precision, recall, F-measure, and area under the receiver operating characteristic curve. External validation was used to examine the stability of the model that showed the best performance.

RESULTS

The logistic model showed the best performance, according to 10-fold cross-validation, among the 6 models, with the highest area under the curve (0.84), accuracy (78.5%), true-positive rate (0.785), sensitivity (0.789), specificity (0.785), precision (0.789), recall (0.785), and F-measure (0.784), as well as the lowest false-positive rate (0.215). The external validation results showed that the logistic model was stable.

CONCLUSIONS

Gray-level co-occurrence matrix texture features extracted from sonograms of thyroid nodules coupled with a logistic model are useful for differentiating between benign and malignant thyroid nodules.