EfficientUNetViT: Efficient Breast Tumor Segmentation Utilizing UNet Architecture and Pretrained Vision Transformer

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Accurate breast tumor segmentation is crucial for diagnosis and treatment planning.
• Existing segmentation models face challenges with computational complexity and generalization.
• Integrating advanced feature extraction with efficient architectures is an ongoing research area.

Purpose of the Study:

• To develop a novel neural network for enhanced breast tumor segmentation.
• To leverage the strengths of Vision Transformers (ViT) and UNet architectures.
• To improve the efficiency and accuracy of medical image segmentation models.

Main Methods:

• A hybrid UNet framework was developed, incorporating a pretrained Vision Transformer (ViT) as the encoder.
• Depthwise separable convolutional blocks were integrated into the UNet to reduce computational load.
• The model was trained and evaluated on medical images for breast tumor segmentation.

Main Results:

• The proposed hybrid model demonstrated superior performance in segmenting breast tumors.
• The integration of ViT enhanced feature extraction and contextual understanding.
• The use of depthwise separable convolutions improved model efficiency and reduced overfitting.

Conclusions:

• The hybrid ViT-UNet model offers a significant advancement in breast tumor segmentation accuracy and efficiency.
• This approach highlights the potential of transformer-based models in medical image processing.
• The study establishes a new benchmark for automated tumor segmentation in clinical applications.