Asymmetric Adaptive Heterogeneous Network for Multi-Modality Medical Image Segmentation.

Area of Science:

• Medical imaging
• Computer vision
• Artificial intelligence

Background:

• Current multi-modality medical image segmentation methods often aggregate data without discrimination.
• Existing approaches overlook the varying contributions of different modalities to visual representation and decision-making.

Purpose of the Study:

• To propose an asymmetric adaptive heterogeneous network for multi-modality image feature extraction with modality discrimination and adaptive fusion.
• To address limitations in current methods by enabling distinct processing and fusion of multi-modality image features.

Main Methods:

• Developed a heterogeneous two-stream asymmetric feature-bridging network for extracting complementary features from auxiliary and leading single-modality images.
• Introduced the Transformer-CNN Feature Alignment and Fusion (T-CFAF) module to enhance leading single-modality information.
• Implemented the Cross-Modality Heterogeneous Graph Fusion (CMHGF) module for adaptive high-level semantic fusion of multi-modality features.

Main Results:

• Demonstrated significant efficiency gains compared to ten existing segmentation models.
• Achieved highly competitive segmentation accuracy across six diverse datasets.
• The proposed asymmetric network effectively handles heterogeneity in multi-modality medical images.

Conclusions:

• The proposed asymmetric adaptive heterogeneous network offers a superior approach to multi-modality medical image segmentation.
• Modality discrimination and adaptive fusion are crucial for maximizing the utility of multi-modal data.
• The method provides a promising direction for advancing medical image analysis and segmentation.