MugenNet: A Novel Combined Convolution Neural Network and Transformer Network with Application in Colonic Polyp Image...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Accurate colonic polyp segmentation is crucial for early detection and diagnosis.
• Convolutional Neural Networks (CNNs) offer segmentation capabilities but suffer from long training times.
• Transformers provide computational efficiency via self-attention but risk information loss.

Purpose of the Study:

• To hybridize CNNs and Transformers to leverage their complementary strengths.
• To develop an efficient and accurate model for colonic polyp image segmentation.
• To introduce MugenNet for enhanced early detection of colonic polyps.

Main Methods:

• Applied the hybridization principle to combine CNN and Transformer architectures.
• Developed and implemented the MugenNet model for colonic polyp image segmentation.
• Conducted comprehensive experiments comparing MugenNet against other CNN models on public datasets.

Main Results:

• MugenNet achieved optimal performance on the ETIS dataset with a mean Dice score of 0.714.
• The model demonstrated a high inference speed of 56 FPS.
• Ablation experiments confirmed the effectiveness of the MugenNet architecture.

Conclusions:

• The proposed hybridization method effectively combines CNN and Transformer advantages.
• MugenNet offers a superior approach for colonic polyp image segmentation.
• This work contributes a computationally efficient and accurate tool for early polyp detection.