Related Experiment Video
Updated: Aug 15, 2025
04:48
Swin-PSAxialNet: An Efficient Multi-Organ Segmentation Technique
Published on: July 5, 2024
478
Fractal, recurrent, and dense U-Net architectures with EfficientNet encoder for medical image segmentation
Nahian Siddique1, Sidike Paheding2, Abel A Reyes Angulo1
1Purdue University Northwest, Department of Electrical and Computer Engineering, Hammond, Indiana, United States.
Journal of Medical Imaging (Bellingham, Wash.)
|January 2, 2023
Summary
This study introduces three novel U-Net variants (efficient R2U-Net, dense U-Net, and fractal U-Net) that improve medical image segmentation accuracy. These models leverage EfficientNet and advanced layer connections to overcome challenges like vanishing gradients in deep learning networks.
Area of Science:
- Deep Learning
- Medical Image Analysis
- Computer Vision
Background:
- U-Net is a foundational deep learning model for medical image segmentation.
- Challenges in deep U-Nets include the black-box problem and vanishing gradients, hindering performance.
- Advanced network designs are crucial for achieving human-like segmentation accuracy.
Purpose of the Study:
- To propose three novel U-Net variants: efficient R2U-Net, efficient dense U-Net, and efficient fractal U-Net.
- To enhance U-Net's capability in creating highly accurate medical image segmentation maps.
- To address the limitations of traditional U-Net architectures, such as vanishing gradients.
Main Methods:
- Developed three U-Net variants by integrating EfficientNet as the encoder.
- Implemented residual connections, dense connections, and fractal expansion in the U-Net decoders.
- Tested models on diverse medical imaging datasets: CHASE DB1, DRIVE, ISIC 2018, and HAM10000.
Main Results:
- Achieved high Dice coefficients: 0.8013 (CHASE DB1), 0.8808 (ISIC 2018), 0.8019 (DRIVE), and 0.9295 (HAM10000).
- Obtained competitive Jaccard (JAC) scores: 0.6686 (CHASE DB1), 0.7870 (ISIC 2018), 0.6694 (DRIVE), and 0.8683 (HAM10000).
- Demonstrated superior segmentation performance compared to existing state-of-the-art models.
Conclusions:
- The proposed U-Net variants show significant improvements in medical image segmentation.
- Integrating EfficientNet and advanced decoder designs enhances U-Net's performance and scalability.
- These adaptable deep learning frameworks offer a promising direction for advanced medical image analysis.
