Frequency-aware Vision Mamba With Deformable Windowed Selective Scan for Ultrasound Image Segmentation

Area of Science:

• Medical Imaging
• Computer Vision
• Biomedical Engineering

Background:

• Accurate segmentation of ultrasound images is vital for medical diagnosis and treatment.
• Challenges include low contrast, high speckle noise, blurred boundaries, and variations in lesion characteristics, leading to inter-class indistinction and intra-class inconsistency.

Purpose of the Study:

• To develop an advanced method for precise lesion segmentation in ultrasound images.
• To overcome the limitations of existing techniques in handling image quality issues and lesion variability.

Main Methods:

• Proposed a frequency-aware vision mamba network incorporating a deformable windowed selective scan.
• Introduced a frequency-aware selective state space layer using Fast Fourier Transform for feature enhancement.
• Developed a deformable windowed selective scan module with a learnable offset field to focus on ambiguous boundaries.

Main Results:

• The frequency-aware vision mamba network demonstrated superior performance across four cross-domain ultrasound datasets.
• Achieved high Dice coefficient scores (0.83-0.87) and Intersection over Union scores (0.76-0.80) in various ultrasound segmentation tasks.
• Outperformed existing state-of-the-art methods in breast, thyroid, and ovarian ultrasound image analysis.

Conclusions:

• The proposed method offers a robust solution for clinical ultrasound image analysis.
• Effectively addresses challenges of inter-class indistinction and intra-class inconsistency in medical image segmentation.