SAU-NET: Scale Aware Polyp Segmentation using Encoder-Decoder Network

Colorectal Cancer has become a major cause of death in recent times. To improve the chances of survival, detecting early signs and identifying polyps in a routine examination is necessary. In this pursuit, an automatic computer-aided diagnosis (CAD) system to detect early signs of the disease onset is crucial. Deep Learning is at the base of recent advances in CAD systems, and its successes in CAD encourage it to be used in colorectal cancer analysis. Efficient segmentation of polyps from the colonoscopy images can aid radiologists immensely in the task of polyps identification and analysis. Therefore, this paper proposes an encoder-decoder-based architecture to segment polyps. Our proposed model takes into account multi-scale features that are present in the images. For efficient feature extraction, residual encoder blocks and Squeeze and Excitation modules are used to enhance channel inter-dependencies. Again, a residual dense decoder is used to improve the reconstruction in the decoder module. During the encoding stage, spatial information loss leads to a semantic gap between encoder and decoder. To handle this, a modified skip connection is proposed to bridge the semantic gap between the encoder and the corresponding decoder. The proposed model achieved an 85.15% dice score, 74.26% IoU, and 83.88% mean IoU on the Kvasir-SEG dataset. The results show that the proposed model can segment polyps more efficiently than many other popular models in the literature. Comparatively, fewer parameters in the proposed model prevail it efficient in real-time use.