Value of artificial intelligence with novel tumor tracking technology in the diagnosis of gastric submucosal tumors by contrast‐enhanced harmonic endoscopic ultrasonography

Abstract

Background and Aim Contrast‐enhanced harmonic endoscopic ultrasonography (CH‐EUS) is useful for the diagnosis of lesions inside and outside the digestive tract. This study evaluated the value of artificial intelligence (AI) in the diagnosis of gastric submucosal tumors by CH‐EUS. Methods This retrospective study included 53 patients with gastrointestinal stromal tumors (GISTs) and leiomyomas, all of whom underwent CH‐EUS between June 2015 and February 2020. A novel technology, SiamMask, was used to track and trim the lesions in CH‐EUS videos. CH‐EUS was evaluated by AI using deep learning involving a residual neural network and leave‐one‐out cross‐validation. The diagnostic accuracy of AI in discriminating between GISTs and leiomyomas was assessed and compared with that of blind reading by two expert endosonographers. Results Of the 53 patients, 42 had GISTs and 11 had leiomyomas. Mean tumor size was 26.4 mm. The consistency rate of the segment range of the tumor image extracted by SiamMask and marked by the endosonographer was 96% with a Dice coefficient. The sensitivity, specificity, and accuracy of AI in diagnosing GIST were 90.5%, 90.9%, and 90.6%, respectively, whereas those of blind reading were 90.5%, 81.8%, and 88.7%, respectively (P = 0.683). The κ coefficient between the two reviewers was 0.713. Conclusions The diagnostic ability of CH‐EUS results evaluated by AI to distinguish between GISTs and leiomyomas was comparable with that of blind reading by expert endosonographers.

Diagnostic accuracy of endoscopic ultrasound with artificial intelligence for gastrointestinal stromal tumors: A meta‐analysis

“…Therefore, subgroup analyses were performed according to the classification of lesions (binary: GIST vs leiomyoma; multicategory: GIST vs non‐GIST). For binary classification (GIST vs leiomyoma), the pooled sensitivity and specificity were 0.93 (95% CI 0.89–0.96; I 2 = 68.9%) and 0.84 (95% CI 0.65–0.93; I 2 = 85.5%), with the AUROC of 0.94 29,31–33 . With respect to multicategory classification (GIST vs non‐GIST), the diagnostic performance of AI‐based EUS for differentiating GISTs remained acceptable (sensitivity 0.90; specificity 0.73; AUROC 0.83) 27,28,30 …”

“…The majority of the included studies were retrospectively designed, though one study evaluated the diagnostic performance of the AI system in both retrospective and prospective forms 32 . Six studies using B‐mode EUS images examined the performance of AI models, whereas the other used an AI model to CH‐EUS images 33 . In terms of validation method, three studies used independent validation, three used internal validation, and one used leave‐one‐out cross‐validation (Table 1).…”

“…In addition, the resolution of EUS images collected from different endoscopic systems may vary across studies, which might have an impact on the diagnostic performance of the AI system being used. Third, we were unable to examine the association between lesion size and diagnostic performance of AI‐based EUS for GISTs owing to a scarcity of studies and data 28,31,33 . Minoda et al 28 reported a better diagnostic efficacy of AI‐based EUS for large GISTs (>20 mm), whereas others observed that the ability of CNN was not affected by tumor size 31,33 .…”

“…It has been shown that the implementation of AI technology in medical areas can improve diagnostic accuracy, facilitate clinical management, and lead to better health outcomes 31 . CNN has been regarded as a reliable model for different visual tasks, resulted in the development of various competitive algorithms, such as AlexNet, GoogleNet, VGG16, YOLO, and Resnet 33 . AI‐based EUS could acquire pixel‐level information, which is impossible with the naked eye.…”

“…In addition, it is worth noting that studies on AI‐based EUS using an image‐enhanced approach remain scarce. Actually, the diagnostic yield of AI‐based CH‐EUS images has only been examined in one study 33 . Subgroup analysis indicated that the sensitivity of conventional EUS (B‐mode) remained similar but the specificity was lower compared to that of CH‐EUS (sensitivity 93%; specificity 77%).…”