A Multi Context Decoder-based Network with Applications for Polyp Segmentation in Colonoscopy Images

Polyp Segmentation is important in helping doctors diagnose and provide an accurate treatment plan. With the emerging of deep learning technology in the last decade, deep learning models especially Unet and its evolved versions, for medical segmentation task have achieved superior results compared to previous traditional methods. To preserve location information, Unet-based models use connections between feature maps of the same resolution of encoder and decoder. However, using the same resolution connections has two problems: 1) High-resolution feature maps on the encoder side contain low-level information. In contrast, high-resolution feature maps on the decoder side contain high-level information that leads to an imbalance in terms of semantic information when connecting. 2) In medical images, objects such as tumours and cells often have diverse sizes, so to be able to segment objects correctly, the use of context information on a scale of the feature map encoder during the decoding process is not enough, so it is necessary to use context information on full-scale. In this paper, we propose a model called CTDCFormer that uses the PvitV2_B3 model as the backbone encoder to extract global information about the object. In order to exploit the full-scale context information of the encoder, we propose the GCF module using the lightweight attention mechanism between the decoder’s feature map and the encoder’s four feature maps. Our model CTDCFormer achieves superior results compared to other state of the arts, with the Dice scores up to 94.1% on the Kvasir-SEG set, and 94.7% on the CVC-ClinicDB set.