MCV-UNet: a modified convolution & transformer hybrid encoder-decoder network with multi-scale information fusion for...

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Accurate semantic segmentation of ultrasound images is crucial for medical diagnosis.
• Deep learning techniques have advanced ultrasound image analysis, but challenges remain in capturing multi-scale features and long-range dependencies.

Purpose of the Study:

• To introduce a novel hybrid network integrating Convolutional Neural Networks (CNN) and Vision Transformers (ViT) for enhanced ultrasound image semantic segmentation.
• To improve feature learning across multiple scales and capture long-range spatial dependencies in ultrasound images.

Main Methods:

• A hybrid network architecture combining multi-scale CNNs in the encoder and decoder with a Vision Transformer (ViT) in the bottleneck.
• Extensive experiments were conducted on a public benchmark ultrasound nerve segmentation dataset.
• The proposed method was benchmarked against 17 existing baseline methods.

Main Results:

• The hybrid CNN-ViT network outperformed all 17 competing methods on the ultrasound nerve segmentation task.
• Achieved a 4.6% improvement in Dice score compared to TransUNet and a 13.0% improvement compared to Attention UNet.
• Demonstrated a 10.5% improvement in precision compared to UNet, highlighting superior performance.

Conclusions:

• The proposed hybrid CNN-ViT network offers a powerful new approach for semantic segmentation in ultrasound imaging.
• Blending CNNs and ViTs in a unified framework significantly enhances feature learning and captures complex spatial relationships.
• This research holds substantial potential for advancing real-world medical imaging applications.