One of the biggest hazards to cancer-related mortality globally is colorectal cancer, and improved patient outcomes are greatly influenced by early identification. Colonoscopy is a highly effective screening method, yet segmentation and detection remain challenging aspects due to the heterogeneity and variability of readers' interpretations of polyps. In this work, we introduce a novel deep learning architecture for gastrointestinal polyp segmentation in the Kvasir-SEG dataset. Our method employs an encoder-decoder structure with a pre-trained ConvNeXt model as the encoder to learn multi-scale feature representations. The feature maps are passed through a ConvNeXt Block and then through a decoder network consisting of three decoder blocks. Our key contribution is the employment of a cross-attention mechanism that creates shortcut connections between the decoder and encoder to maximize feature retention and reduce information loss. In addition, we introduce a Residual Transformer Block in the decoder that learns long-term dependency by using self-attention mechanisms and enhance feature representations. We evaluate our model on the Kvasir-SEG dataset, achieving a Dice coefficient of 0.8715 and mean intersection over union (mIoU) of 0.8021. Our methodology demonstrates state-of-the-art performance in gastrointestinal polyp segmentation and its feasibility of being used as part of clinical pipelines to assist with automated detection and diagnosis of polyps.
Keywords: Deep learning; Gastrointestinal polyp; Image segmentation; Kvasir-SEG; Transformer.
© 2025 Wang et al.