A multi-branched semantic segmentation network based on twisted information sharing pattern for medical images

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence in Healthcare

Background:

• Semantic segmentation is crucial for smart healthcare applications like diagnosis and treatment planning.
• Current medical image segmentation faces accuracy challenges due to poor feature interactivity and limited local feature exploration.

Purpose of the Study:

• To introduce a novel architecture, TP-MNet, to enhance feature fusion and improve medical image semantic segmentation accuracy.
• To address the limitations of semantic isolation and inadequate local feature mining in existing models.

Main Methods:

• Proposed the Twisted Information-sharing Pattern for Multi-branched Network (TP-MNet) architecture.
• Implemented mutual feature transfer between neighboring branches for semantic fusion.
• Incorporated secondary feature mining during transfer and refined feature fusion modules for contextual information acquisition.

Main Results:

• TP-MNet demonstrated superior performance across 5 medical datasets compared to 21 other models.
• Extensive validation through metric analysis, image comparisons, and ablation tests confirmed the model's effectiveness.
• Investigations clarified the practical utility and limitations of the Twisted Information-sharing Pattern.

Conclusions:

• The Twisted Information-sharing Pattern significantly improves semantic fusion, enhancing medical image segmentation.
• This semantic broadcasting approach highlights the importance of semantic fusion and advances multi-branched architectures in medical AI.