The integration of Artificial Intelligence (AI), especially deep learning, is of paramount importance in reducing mortality linked to gastrointestinal diseases and in the early detection of gastrointestinal cancer. However, such systems still suffer of various problems namely variation of luminosity, computational cost, and lucks of dataset, which attract a lot of researches. Motivated by the mentioned challenges, we propose in the present study a new system for automatic gastrointestinal cancer detection. The proposed solution is composed of three main stages: firstly, a new approach based Encoder-Decoder Network (EDN) architecture was proposed for detection the region of interest (ROI) from an endoscopy image; then, a deep convolutional neural network (CNN) was used to extract descriptive vectors; thereafter, a Support Vector Machine (SVM) classifier was used to recognize normal individuals from affected. The experimental results demonstrate that the proposed approach archives good results with 94% using the kvasir Dataset.