Related Experiment Video
Updated: May 12, 2026
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
3.3K
Enhanced multi-scale selective attention U-net for breast ultrasound image segmentation
Peng Wang1, Keyu Chen2, Shasha Hong1
1School of Electronic Information, Hunan First Normal University, Changsha, China.
Medical Physics
|November 27, 2025
Summary
This study introduces an enhanced multi-scale selective attention (EMSA) U-Net for improved breast ultrasound image segmentation. The model significantly boosts accuracy in identifying lesions, aiding clinical diagnosis.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Computer Vision
Background:
- Accurate segmentation of breast lesions in ultrasound images is critical for diagnosis and treatment.
- Challenges include noise, low contrast, and diverse lesion characteristics.
Purpose of the Study:
- To develop an efficient segmentation model, the enhanced multi-scale selective attention (EMSA) U-Net.
- To improve feature representation using multi-scale context and enhance boundary detection with adaptive attention.
Main Methods:
- Utilized Breast Ultrasound Images (BUSI) and BUSI_WHU datasets.
- Employed data augmentation and preprocessing for model training.
- Integrated EMSA blocks with multi-branch dilated convolution and SKConv for adaptive feature fusion.
Main Results:
- Achieved an mDice score of 68.33% on the BUSI test set.
- Demonstrated statistically significant improvements over TransUNet and U-Net.
- Showed a large effect size compared to CSA-UNet.
Conclusions:
- The EMSA U-Net model offers superior segmentation performance for breast ultrasound images.
- The model effectively integrates adaptive multi-scale feature learning and context-aware attention.
