ESAM2-BLS: Enhanced segment anything model 2 for efficient breast lesion segmentation in ultrasound imaging

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Ultrasound imaging is crucial for breast lesion diagnosis but faces segmentation challenges due to noise and image quality variations.
• Accurate segmentation of breast lesions is vital for effective screening and diagnosis.

Purpose of the Study:

• To develop a novel deep learning model, ESAM2-BLS, for improved breast lesion segmentation in ultrasound images.
• To address specific challenges in breast ultrasound imaging, including speckle noise and low contrast.

Main Methods:

• An enhanced Segment Anything Model 2 (SAM2) architecture, ESAM2-BLS, was developed using an adapter module tailored for breast ultrasound images.
• The model incorporates channel attention, specialized convolutions, and optimized skip connections to handle ultrasound-specific artifacts.
• Multi-scale feature fusion and axial dilated depthwise convolution were employed for comprehensive lesion information capture.

Main Results:

• ESAM2-BLS demonstrated significantly improved segmentation accuracy, especially for low-contrast and small lesions.
• The model achieved an average Dice score of 0.9077 and 0.8633 in five-fold cross-validation across two datasets.
• Robustness and accuracy improvements were observed compared to traditional segmentation methods.

Conclusions:

• ESAM2-BLS offers an efficient, reliable, and specialized automated solution for breast lesion segmentation.
• The model enhances early breast cancer screening and diagnosis by improving segmentation performance.
• The customized adapter module effectively tackles unique challenges in breast ultrasound imaging.