SC-Net: Symmetrical conical network for colorectal pathology image segmentation.

Area of Science:

• Medical Image Analysis
• Computer-Aided Diagnosis
• Histopathology Segmentation

Background:

• Convolutional neural networks struggle with extracting heterogeneous semantic information and context dependency across different receptive fields in histopathology images.
• Existing linear flow structures in neural networks limit the ability to capture multi-level feature variations crucial for accurate medical image diagnosis.

Purpose of the Study:

• To propose a novel Symmetric Conical Network (SC-Net) for enhanced image segmentation of colorectal cancer histopathology.
• To address limitations in extracting multi-scale, heterogeneous semantic information and establishing context dependency in existing deep learning models.

Main Methods:

• Developed a Symmetric Conical Network (SC-Net) featuring a Multi-scale Feature Extraction Block (MFEB) for diverse feature extraction.
• Employed a spiral and multi-branch arrangement of MFEBs in the encoder to improve context dependence among information flows.
• Introduced a Feature Mapping Layer (FML) to map low-level to high-level semantic features, mitigating information loss and enhancing global feature extraction.

Main Results:

• SC-Net achieved segmentation mDice scores of 0.8611 on colorectal cancer, 0.7259 on breast cancer, and 0.7144 on polyp datasets.
• Outperformed state-of-the-art models including UNet++, PSPNet, Attention U-Net, and R2U-Net in segmentation tasks.
• Demonstrated superior performance in segmenting H&E stained pathology images across various datasets.

Conclusions:

• The SC-Net effectively segments H&E stained pathology images, preserving crucial morphological and spatial information.
• The network shows robustness in handling challenging image conditions such as weak texture, poor contrast, and appearance variations.
• SC-Net represents a significant advancement in computer-aided medical image diagnosis for colorectal cancer and other pathologies.