Multi-granularity cross-image semantic modeling for medical ultrasound image segmentation

Area of Science:

• Medical Imaging
• Computer-Aided Diagnosis
• Artificial Intelligence

Background:

• Automatic segmentation of medical ultrasound images is crucial for computer-aided diagnosis (CAD).
• Existing methods often neglect cross-image semantic modeling, limiting robustness against variations.
• Over-reliance on global context can overlook fine-grained lesion details.

Purpose of the Study:

• To propose Multi-granularity Cross-image Semantic Modeling (MCSM) for improved ultrasound image segmentation.
• To address limitations of existing methods by incorporating both global and local representations.
• To enhance segmentation reliability by mitigating semantic conflicts and noise.

Main Methods:

• Developed a Multi-Granularity Patch Extraction Process (MGPEP) to fuse whole-image and lesion-centric features.
• Introduced a Frequency-domain Feature Dependency Module (FFDM) for multi-granularity spectral filtering and cross-image dependency modeling.
• Integrated preceding batch features into current batch representations in the frequency domain to improve discrimination.

Main Results:

• MCSM demonstrated superior effectiveness and robustness compared to state-of-the-art segmentation methods.
• The proposed method successfully enhances foreground-background discrimination.
• Experimental results highlight the potential of MCSM for clinical applications.

Conclusions:

• MCSM effectively models cross-image semantics at multiple granularities, improving segmentation accuracy.
• The method addresses challenges like inter-patient variations and noise in ultrasound imaging.
• MCSM shows significant promise for advancing computer-aided diagnosis in medical imaging.