An embedded recurrent neural network-based model for endoscopic semantic segmentation

Detecting cancers at their early stage would decrease mortality rate. For instance, detecting all polyps during colonoscopy would increase the chances of a better prognoses. However, endoscopists are facing difficulties due to the heavy workload of analyzing endoscopic images. Hence, assisting endoscopist while screening would decrease polyp miss rate. In this study, we propose a new deep learning segmentation model to segment polyps found in endoscopic images extracted during Colonoscopy screening. The propose model modifies SegNet architecture to embed Gated recurrent units (GRU) units within the convolution layers to collect contextual information. Therefore, both global and local information are extracted and propagated through the entire layers. This has led to better segmentation performance compared to that of using state of the art SegNet. Four experiments were conducted and the proposed model achieved a better intersection over union “IoU” by 1.36%, 1.71%, and 1.47% on validation sets and 0.24% on a test set, compared to the state of the art SegNet.