TAC-UNet: transformer-assisted convolutional neural network for medical image segmentation.

Area of Science:

• Medical image analysis
• Artificial intelligence in healthcare
• Deep learning for segmentation

Background:

• Convolutional Neural Networks (CNNs) are limited in capturing global context in medical images.
• Vision Transformers (ViTs) capture global context but require large datasets, which are scarce in medical imaging.
• Integrating CNNs and ViTs can leverage their complementary strengths for improved medical image segmentation.

Purpose of the Study:

• To develop a novel hybrid deep learning architecture for medical image segmentation.
• To address the limitations of CNNs and ViTs in small-scale medical image datasets.
• To improve the accuracy and efficiency of medical image segmentation.

Main Methods:

• A U-shaped network architecture named TAC-UNet was developed, integrating CNN and Transformer components.
• A dual-path design allows the Transformer branch to provide global contextual information to the CNN backbone.
• A channel cross-attention (CCA) module was incorporated to reconcile semantic discrepancies between the encoder and decoder.

Main Results:

• TAC-UNet achieved state-of-the-art Dice scores on three public datasets: MoNuSeg (80.36%), GlaS (90.70%), and CVC-ClinicDB (91.81%).
• The hybrid model demonstrated significantly superior segmentation performance compared to CNN-based, Transformer-based, and other hybrid methods.
• Experiments confirmed the effectiveness of the TAC-UNet architecture for small-scale medical image segmentation.

Conclusions:

• The TAC-UNet model offers advanced segmentation performance for small-scale medical image datasets.
• The hybrid approach effectively combines local feature extraction of CNNs with global context modeling of Transformers.
• The developed model shows significant promise for enhancing healthcare outcomes through improved medical image analysis.