Automatically optimized radiomics modeling system for small gastric submucosal tumor (<2 cm) discrimination based on EUS images

Background and Aims The clinical management of small gastric submucosal tumors (SMTs) (<2cm) faces a non-negligible challenge due to the lack of guideline consensus and effective diagnostic tools. This paper develops an automatically optimized radiomics modeling system (AORMS) based on endoscopic ultrasound (EUS) images to diagnose and evaluate SMTs. Methods EUS images of 205 small gastric SMT (<2cm) patients were retrospectively enrolled in the development phase of AORMS, for the diagnosis and the risk stratification of gastrointestinal stromal tumor (GIST). Images of 178 patients from different centers were prospectively enrolled in the independent testing phase. The performance of the AORMS was compared to that of endoscopists in the development set and evaluated in the independent testing set. Results The AORMS demonstrated an area under the curve (AUC) of 0.762 for the diagnosis of GIST, while 0.734 for the risk stratification of GIST, respectively. In the independent testing set, the AORMS achieved an AUC of 0.770 and 0.750 for the diagnosis and risk stratification of small GISTs, respectively. In comparison, the AUC of five experienced endoscopists ranged from 0.501-0.608 for diagnosing GIST, and 0.562-0.748 for risk stratification. The AORMS outperformed experienced endoscopists by over 20% in diagnosing GIST. Conclusions The AORMS implements automatic parameter selection, which enhances its robustness and clinical applicability. It has demonstrated good performance in the diagnosis and risk stratification of GISTs, which could aid endoscopists in the diagnosis of small gastric SMTs (<2cm).