A combined radiomic model distinguishing GISTs from leiomyomas and schwannomas in the stomach based on endoscopic ultrasonography images

Abstract

Background Endoscopic ultrasonography (EUS) is recommended as the best tool for evaluating gastric subepithelial lesions (SELs); nonetheless, it has difficulty distinguishing gastrointestinal stromal tumors (GISTs) from leiomyomas and schwannomas. GISTs have malignant potential, whereas leiomyomas and schwannomas are considered benign. Purpose This study aimed to establish a combined radiomic model based on EUS images for distinguishing GISTs from leiomyomas and schwannomas in the stomach. Methods EUS images of pathologically confirmed GISTs, leiomyomas, and schwannomas were collected from five centers. Gastric SELs were divided into training and testing datasets based on random split‐sample method (7:3). Radiomic features were extracted from the tumor and muscularis propria regions. Principal component analysis, least absolute shrinkage, and selection operator were used for feature selection. Support vector machine was used to construct radiomic models. Two radiomic models were built: the conventional radiomic model included tumor features alone, whereas the combined radiomic model incorporated features from the tumor and muscularis propria regions. Results A total of 3933 EUS images from 485 cases were included. For the differential diagnosis of GISTs from leiomyomas and schwannomas, the accuracy, sensitivity, specificity, and area under the receiver operating characteristic curve were 74.5%, 72.2%, 78.7%, and 0.754, respectively, for the EUS experts; 76.8%, 74.4%, 81.0%, and 0.830, respectively, for the conventional radiomic model; and 90.9%, 91.0%, 90.6%, and 0.953, respectively, for the combined radiomic model. For gastric SELs <20 mm, the accuracy, sensitivity, specificity, and area under the receiver operating characteristic curve of the combined radiomic model were 91.4%, 91.6%, 91.1%, and 0.960, respectively. Conclusions We developed and validated a combined radiomic model to distinguish gastric GISTs from leiomyomas and schwannomas. The combined radiomic model showed better diagnostic performance than the conventional radiomic model and could assist EUS experts in non‐invasively diagnosing gastric SELs, particularly gastric SELs <20 mm.

Endoscopic ultrasonography-based intratumoral and peritumoral machine learning radiomics analyses for distinguishing insulinomas from non-functional pancreatic neuroendocrine tumors

“…Consistently, prior research has demonstrated that the integration of peritumoral and intratumoral data using a nomogram model, which incorporates deep learning contrast-enhanced ultrasound and clinical characteristics, has exhibited notable proficiency in the identification of preoperative aggressiveness in PNETs ( 33 ). Moreover, the effectiveness of employing radiomics, machine learning, and deep learning techniques based on EUS imaging for the prediction of gastrointestinal stromal tumors and pancreatic ductal adenocarcinoma has been substantiated in previous studies ( 30 , 39 , 40 ). However, to the best of our knowledge, we were the first to report on the remarkable predictive capabilities of EUS imaging-based intratumoral and/or peritumoral radiomics models for identifying NF-PNETs and insulinomas.…”

Construction and validation of an endoscopic ultrasonography-based ultrasomics nomogram for differentiating pancreatic neuroendocrine tumors from pancreatic cancer

“…This model effectively mitigated the diagnostic disparity among ultrasound endoscopy physicians of varying expertise levels, thereby enhancing the precision of their diagnoses (31). The EUS-based ultrasomics model was developed to accurately distinguish between gastric GISTs, smooth muscle tumors, and nerve sheath tumors (32). Regrettably, the literature lacked published studies that have utilized EUS imaging ultrasomics to diagnose and predict PNET.…”