AI-assisted polyps detection to facilitate autonomous endoscopy examination

Examining the gastrointestinal (GI) tract to find polyps is a major task in endoscopy to diagnose, monitor and treat digestive disorders and diseases. Traditionally, an expert performs this examination to identify polyps based on the endoscope-generated images displayed on a screen in real-time. However, this manual identification is time consuming, tedious, suffers from inter-observer variability and often leads to misidentification. An automated method would be more effective; however, existing methods fail to achieve adequate accuracy and usability. This paper presents a two-step polyps detection method that classifies polyps images first and then segments polyps from them, suitable for practical use. The proposed method utilized a transformer for classification and a residual UNet model for segmentation. This method achieved 99.9% accuracy for classifying polyps images and 98.5% accuracy in segmenting the polyps when demonstrated on the HyperKvasir dataset, outperforming the state-of-the-art. HyperKvasir is the most commonly used public dataset for GI landmarks and anomalies detection. Effective design and high accuracy ensured the practical use of the system.