Cross-Parallel Transformer: Parallel ViT for Medical Image Segmentation.

Area of Science:

• Medical Image Analysis
• Computer Vision
• Artificial Intelligence

Background:

• Medical image segmentation commonly uses hybrid models combining Convolutional Neural Networks and sequential Transformers.
• Transformers employ multi-head self-attention for global context but suffer from inefficient feature extraction and high computational demands, impacting robustness.

Purpose of the Study:

• To develop more efficient and robust medical image segmentation models.
• To address the computational inefficiencies of current Transformer-based approaches in medical imaging.

Main Methods:

• Introduced PTransUNet (PT model) and C-PTransUNet (C-PT model) for medical image segmentation.
• The C-PT module modifies the Vision Transformer architecture from sequential to parallel processing.
• Enhanced Multi-Head Self-Attention with self-correlated feature attention and channel feature interaction, while optimizing the Feed-Forward Network.

Main Results:

• The C-PTransUNet model achieved a 3.25% improvement in Dice Similarity Coefficient (DSC) accuracy on the Synapse dataset compared to the baseline.
• The PT model showed comparable performance to the baseline in terms of parameters and FLOPs.
• The C-PT model reduced parameter count by 29% and FLOPs by 21.4% compared to the baseline.

Conclusions:

• The proposed PTransUNet and C-PTransUNet models offer significant improvements in both segmentation accuracy and computational efficiency.
• C-PTransUNet presents a particularly effective solution for enhancing Transformer-based medical image segmentation performance while reducing resource requirements.