PLPdra: Pyramid Linear Prediction decoder and reverse attention for Colon Polyps Segmentation

Zhaolin Lu¹Yu Tian²Hongliang Bi²Wei Wang¹Bohe Wang¹Guangxia Chen¹Shiyu Liu^1*

• ¹Xuzhou Municipal First People's Hospital, Xuzhou, China
• ²China University of Mining and Technology, Xuzhou, China

Artificial intelligence (AI) methods are increasingly used to assist in diagnosis and treatment, particularly for the detection of colon polyps during colonoscopy. However, the detection accuracy of colon polyps is significantly impacted by missed detection due to small polyp targets and edge blurring resulting from the complex intestinal environment. To address these challenges, we propose a novel medical computer-aided diagnosis algorithm, called Pyramid Linear Prediction Decoder Reverse Attention (PLPdra). The model improves the segmentation accuracy for small polyp targets and enhances the clarity of the segmentation boundaries. PLPdra follows an encoder-decoder architecture. The encoder utilizes the Pyramid Vision Transformer, which extracts both local and global features more effectively than traditional CNNs or Transformers. A Linear-Stage Feature Predictive Decoder (LSFPD) module is introduced between the encoder and decoder to facilitate dual localization of small polyp targets. Additionally, a dual-channel Reverse Axial Target Attention (RATA) module is integrated into the decoder to address the issue of edge blurring in segmentation. We evaluated PLPdra on five benchmark datasets: Kvasir, CVC-ClinicDB, CVC-ColonDB, CVC-300, and ETIS-Larib. Specifically, we conducted targeted experiments on the small polyp targets and edge blurring categories within the aforementioned datasets. The experimental results demonstrate that PLPdra achieves competitive performance in terms of mDice segmentation accuracy, improving the detection rate of colon polyps.