DMAeEDNet: Dense Multiplicative Attention Enhanced Encoder Decoder Network for Ultrasound-Based Automated Breast...

Area of Science:

• Medical imaging
• Artificial intelligence
• Biomedical engineering

Background:

• Accurate breast lesion segmentation in ultrasound images is crucial for early breast cancer diagnosis.
• Deep learning models, particularly U-Net, are widely used but often increase computational time.
• There is a need for efficient segmentation networks that maintain high accuracy.

Purpose of the Study:

• To develop a low-complexity deep learning network for effective breast lesion segmentation in ultrasound images.
• To enhance feature selectivity using attention mechanisms within an encoder-decoder architecture.
• To achieve real-time segmentation performance without compromising accuracy.

Main Methods:

• Proposed a dense multiplicative attention enhanced encoder-decoder network.
• Integrated two dense multiplicative attention components in encoding and output layers.
• Utilized depthwise separable convolutions for feature enhancement.
• Evaluated performance on two public datasets and an in-clinic dataset.

Main Results:

• Achieved Dice coefficients of 0.83 and 0.86 on public datasets.
• Demonstrated an average segmentation latency of 19ms.
• Obtained a Dice coefficient of 0.72 on a noise-robust in-clinic dataset.
• Showcased superior performance compared to commonly used networks in terms of speed and complexity.

Conclusions:

• The proposed network offers effective and efficient breast lesion segmentation in ultrasound images.
• The attention mechanisms enhance relevant features, leading to improved segmentation accuracy.
• The low computational complexity and high speed make the network suitable for real-time clinical applications.
• This approach presents a feasible solution for improving early breast cancer diagnosis through automated image analysis.