Semantic Segmentation of Colorectal Polyps with DeepLab and LSTM Networks

In this work we attempted to use the existing deep neural network called DeepLab_v3 to detect the polyps in colonoscopy images. Due to its large structure, the location of polyps may not be preserved and transmitted effectively. To address the issue we combined Long Short-Term Memory networks and DeepLab_v3 in parallel to augment the signal of the polyps' location. The new modification was examined with the colonoscopy image database `CVC-ClinicDB' from MICCAI sub-challenge 2015. After training with 267 images and testing with 345 images, we got a good performance, 93.21% mean Intersection over Union (mIOU). The average computing time is 0.023 second per image. Once the model is applied to clinical colonoscopy exam videos, it could provide effective second opinions in real time to aid the diagnosis.