Related Experiment Video
Updated: May 6, 2026
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
3.6K
Inception U-Net for Enhanced Breast Ultrasound Image Segmentation Using Transfer Learning
Yeonhyo Choi1, Myoung Nam Kim2, Sungdae Na3
1Department of Medical & Biological Engineering, Graduate School, Kyungpook National University, Daegu 41404, Republic of Korea.
Bioengineering (Basel, Switzerland)
|February 27, 2026
Summary
This study enhances breast cancer segmentation in ultrasound images by integrating Inception modules into the U-Net architecture. The improved model shows a ~5% performance boost, advancing automated medical image analysis.
Area of Science:
- Medical Imaging
- Artificial Intelligence in Medicine
- Image Segmentation
Background:
- Breast cancer diagnosis relies heavily on ultrasound imaging.
- Operator dependency and image quality issues hinder traditional methods.
- Existing U-Net models have limitations in feature extraction due to shallow encoders.
Purpose of the Study:
- To develop an enhanced segmentation model for breast ultrasound images.
- To improve feature extraction capabilities beyond traditional U-Net architectures.
- To leverage transfer learning for better segmentation performance.
Main Methods:
- Replaced the U-Net encoder with an Inception architecture.
- Utilized transfer learning with ImageNet pre-trained weights.
- Trained and evaluated the model on 900 breast ultrasound images.
Main Results:
- The Inception U-Net achieved an IoU of 0.7774 and Dice score of 0.8491.
- Demonstrated approximately 5% improvement over the baseline U-Net.
- Achieved precision of 0.7081 and recall of 0.7174.
Conclusions:
- Inception modules enhance feature extraction for breast ultrasound segmentation.
- Transfer learning from ImageNet is effective despite domain differences.
- The approach provides a foundation for advanced medical imaging applications.
