HSNet

HSNet improves polyp segmentation in clinical colonoscopy images by combining Transformer architectures and convolutional neural networks to capture long-range dependencies and preserve local appearance details across variations in scale, orientation, and illumination.

Key Features:

• Transformer–CNN hybrid architecture: HSNet integrates Transformer architectures and convolutional neural networks to leverage long-range dependencies alongside local feature extraction.
• Cross-Semantic Attention (CSA) module: CSA bridges low-level and high-level features by exchanging semantic information between different types of network attentions to enhance feature representation.
• Hybrid Semantic Complementary (HSC) module: HSC uses a dual-branch structure combining Transformer and CNN elements to capture both long-range dependencies and local appearance details.
• Multi-Scale Prediction (MSP) module: MSP learns to fuse stage-level prediction masks from the decoder to produce the final segmentation output.
• Robustness to polyp variability: The architecture addresses variability in polyp scale, orientation, and illumination and preserves fine textures to improve segmentation of smaller polyps.
• Experimental evaluation: HSNet was evaluated against 10 state-of-the-art methods on Kvasir-SEG, ClinicDB, ColonDB, ETIS, and Endoscene using mean Dice coefficient (mDic) and mean Intersection over Union (mIoU).

Scientific Applications:

• Polyp segmentation in clinical colonoscopy images: Accurate delineation of polyp regions to assist identification and analysis during colorectal screening.
• Detection of small or low-contrast polyps: Improved segmentation of small and low-contrast lesions through preservation of fine textures and multi-scale fusion.
• Benchmarking segmentation methods: Comparative evaluation of segmentation performance using public datasets Kvasir-SEG, ClinicDB, ColonDB, ETIS, and Endoscene with mDic and mIoU metrics.

Methodology:

Hybrid model combining Transformer and convolutional neural network components; Cross-Semantic Attention (CSA) exchanges semantic information between low- and high-level features; Hybrid Semantic Complementary (HSC) implements a dual-branch Transformer+CNN design; Multi-Scale Prediction (MSP) fuses decoder stage prediction masks; evaluated using mean Dice coefficient (mDic) and mean Intersection over Union (mIoU) on Kvasir-SEG, ClinicDB, ColonDB, ETIS, and Endoscene against 10 state-of-the-art methods.