A Medical Image Segmentation Method Based on Improved UNet 3+ Network.

Area of Science:

• Medical Image Analysis
• Deep Learning
• Computer-Aided Diagnosis

Background:

• Medical image segmentation is crucial for clinical applications, demanding high accuracy and computational efficiency.
• UNet and its variant UNet 3+ are widely used deep learning models, but UNet 3+'s full-scale skip connections can lead to redundant computations.
• There is a need to reduce network parameters and enhance feature extraction in UNet-based models for improved performance.

Purpose of the Study:

• To reduce network parameters and computational redundancy in UNet 3+.
• To improve the feature extraction capability and segmentation accuracy of UNet 3+.
• To validate the proposed model's effectiveness across diverse medical imaging datasets.

Main Methods:

• Pruning full-scale skip connections in UNet 3+ to enhance computational efficiency.
• Integrating the Convolutional Block Attention Module (CBAM) to capture essential features and improve feature expression.
• Evaluating the modified UNet 3+ model on skin cancer, breast cancer, and lung segmentation datasets.

Main Results:

• The proposed model achieved approximately 36% and 18% parameter reduction compared to UNet and UNet 3+, respectively.
• Demonstrated superior performance across various evaluation metrics compared to existing models.
• Achieved more accurate medical image segmentation results on tested datasets.

Conclusions:

• The optimized UNet 3+ model offers efficient feature extraction and improved segmentation performance with fewer parameters.
• The model shows significant potential for enhancing computer-aided diagnosis in medical imaging.
• This approach effectively balances segmentation accuracy and computational efficiency for clinical applications.