OBJECTIVE: The purpose of this study was to explore the role of digital image analysis in differentiating endoscopic ultrasonographic (EUS) features of potentially malignant gastrointestinal subepithelial lesions (SELs) from those of benign lesions. METHODS: Forty-six patients with histopathologically confirmed gastrointestinal stromal tumors (GISTs), carcinoids, and lipomas who had undergone EUS evaluation were identified from our database. Representative regions of interest (ROIs) were selected from the EUS images, and features were extracted by texture analysis. On the basis of these features, an artificial neural network (ANN) was built, trained, and internally validated by unsupervised learning followed by supervised learning. Outcomes were the performance characteristics of the ANN. RESULTS: A total of 106, 111, and 124 ROIs were selected from EUS images of 8, 10, and 28 patients with lipomas, carcinoids, and GISTs, respectively. For each ROI, 228 statistical parameters were extracted and later reduced to the 11 most informative features by principal component analysis. After training with 50% of the data, the remainder of the data were used to validate the ANN. The model was "good" in differentiating carcinoids and GISTs, with area under the receiver operating characteristic curve (AUC) values of 0.86 and 0.89, respectively. The model was "excellent" in identifying lipomas correctly, with an AUC of 0.92. CONCLUSIONS: Digital image analysis of EUS images is a useful noninvasive adjunct to EUS evaluation of SELs.
OBJECTIVE: The purpose of this study was to explore the role of digital image analysis in differentiating endoscopic ultrasonographic (EUS) features of potentially malignant gastrointestinal subepithelial lesions (SELs) from those of benign lesions. METHODS: Forty-six patients with histopathologically confirmed gastrointestinal stromal tumors (GISTs), carcinoids, and lipomas who had undergone EUS evaluation were identified from our database. Representative regions of interest (ROIs) were selected from the EUS images, and features were extracted by texture analysis. On the basis of these features, an artificial neural network (ANN) was built, trained, and internally validated by unsupervised learning followed by supervised learning. Outcomes were the performance characteristics of the ANN. RESULTS: A total of 106, 111, and 124 ROIs were selected from EUS images of 8, 10, and 28 patients with lipomas, carcinoids, and GISTs, respectively. For each ROI, 228 statistical parameters were extracted and later reduced to the 11 most informative features by principal component analysis. After training with 50% of the data, the remainder of the data were used to validate the ANN. The model was "good" in differentiating carcinoids and GISTs, with area under the receiver operating characteristic curve (AUC) values of 0.86 and 0.89, respectively. The model was "excellent" in identifying lipomas correctly, with an AUC of 0.92. CONCLUSIONS: Digital image analysis of EUS images is a useful noninvasive adjunct to EUS evaluation of SELs.