Real-time deployment and test set analysis of automatic colonoscopy polyp identification architecture

Manual detection of polyps in the colonoscopy videos can be challenging due to the large no. of sequential frames, and high chances of false positive rate. So, researchers are working to develop an automatic computer-aided decision system to successfully identify polyps in real-time. The present work proposes a real-time deployment of automatic polyp identification based VGG-16 architecture using flutter application (app) called the Automatic Colonoscopy Polyp Identification (ACPI) testing app. A test set analysis of three open source datasets namely CVC Clinic, CVC Colon, CVC 300 and one private dataset has been done to check the efficacy of the proposed app using various evaluation metrics such as accuracy, precision, recall and F1 score. The proposed app has achieved an average recall, and F1 score up-to 0.7956 and 0.7219 on the developed private dataset consisting of varied colonoscopy polyp and non-polyp frames. The proposed app is immune to inter and intra-device variation and has a user friendly interface. It has achieved encouraging results in comparison to the existing polyp detector app available on Google store, further making it suitable for real-time deployment in this research field.