Application of Computer-Assisted Endoscopic Ultrasonography Based on Texture Features in Differentiating...

Gastrointestinal stromal tumors are common gastric mesenchymal tumors that are potentially malignant. However, endoscopic ultrasonography is poor in diagnosing gastrointestinal stromal tumors. The study investigated the efficacy of texture features extracted from endoscopic ultrasonography images to differentiate gastrointestinal stromal tumors from gastric mesenchymal tumors.

The endoscopic ultrasonography examinations of 120 patients with confirmed gastric gastrointestinal stromal tumors, leiomyoma, or schwannoma were evaluated. Histology was considered the gold standard. Three feature combinations were extracted from endoscopic ultrasonography images of each lesion: 48 gray-level co-occurrence matrix-based features, 48 gray-level co-occurrence matrix-based features plus 3 global gray features, and 15 gray-gradient co-occurrence matrix-based features. Support vector machine classifiers were constructed by using feature combinations to diagnose gastric gastrointestinal stromal tumors. The area under the receiver operating characteristic curve, accuracy, sensitivity, and specificity were used to evaluate the diagnostic performance. The support vector machine model's diagnostic performance was compared with the endoscopists.

The 3 feature combinations had better performance in differentiating gastrointestinal stromal tumors: gray-gradient cooccurrence matrix-based features yielded an area under the receiver operating characteristic curve of 0.90, which was significantly greater than an area under the receiver operating characteristic curve of 0.83 in gray-level co-occurrence matrix-based features and an area under the receiver operating characteristic curve of 0.84 in the texture features plus 3 global features. The support vector machine model (81.67% accuracy, 81.36% sensitivity, and 81.97% specificity) was also better than endoscopists (an average of 69.31% accuracy, 65.54% sensitivity, and 72.95% specificity) Conclusion: Texture features in computer-assisted endoscopic ultrasonography diagnosis are useful to differentiate gastrointestinal stromal tumors from benign gastric mesenchymal tumors and compare favorably with endoscopists. Support vector machine model using gray-gradient co-occurrence matrix-based texture features revealed the best diagnostic performance in diagnosing gastric gastrointestinal stromal tumors.