An artificial intelligence (AI) system called “ENDOANGEL” was effective for real-time monitoring of endoscopic “blind spots” and improved detection of early gastric cancer (EGC) during esophagogastroduodenoscopy (EGD), according to research published recently.
While EGD is widely used to examine lesions found in the upper gastrointestinal tract, there is considerable variability among endoscopists regarding performance, resulting in a substantial miss rate for EGC. But in a study published in the journal, researchers suggest a more objective assessment of lesions with AI technology could improve detection rates in real time, thus improving the chances of establishing an early diagnosis and initiating prompt treatment of gastric cancer.
The researchers updated a developed AI system called WISENSE, which previously demonstrated an ability to monitor gastric areas overlooked during EGD (termed “blind spots”). The investigators integrated a trained real-time EGC detection model into the WISENSE system and changed the name of the updated system to ENDOANGEL.
Researchers from the Renmin Hospital of Wuhan (China) University used deep convolutional neural networks and deep reinforcement learning to develop the ENDOANGEL. A total of 1,050 patients from five hospitals in China who were undergoing EGD were randomized to either an ENDOANGEL-assisted protocol (n = 498) or a control group (n = 504) that did not use the ENDOANGEL system. Examination consisted of white-light imaging observation, magnifying image-enhanced endoscopy observation, and biopsy of suspicious lesions.
The investigators compared the groups in terms of the number of blind spots after the intervention. They assessed the performance of the AI-based ENDOANGEL system in its ability to predict EGC in a real-world clinical setting.
Patients assigned to ENDOANGEL had a significantly fewer mean number of blind spots compared with patients assigned to control (5.38 vs. 9.82, respectively; P < .001). Despite this advantage, patients in the ENDOANGEL group had significantly longer inspection time (5.40 minutes vs. 4.38 minutes; P < .001).
There were 819 lesions reported by endoscopists in the ENDOANGEL group, which included 196 gastric lesions with pathological results. According to the investigators, the ENDOANGEL system correctly predicted all three EGCs, including one mucosal carcinoma and two high grade neoplasias, as well as two advanced gastric cancers. The per-lesion accuracy was 84.7 %, while the sensitivity and specificity rates for detecting gastric cancer were 100% and 84.3%, respectively.
The authors noted limitations of the analysis itself and those stemming from the short follow-up, as well as possible bias introduced by unblinded statisticians. Further research is warranted, they wrote.
“In conclusion, ENDOANGEL, a system for improving endoscopy quality based on deep learning, achieved real-time monitoring of endoscopic blind spots, timing, and EGC detection during EGD,” according to the authors. “ENDOANGEL greatly improved the quality of EGD in this multicenter study, and showed potential for detecting EGC in real clinical settings.”