Alternate encoder and dual decoder CNN-Transformer networks for medical image segmentation

Area of Science:

• Medical Image Analysis
• Computer Vision
• Deep Learning

Background:

• Accurate medical image segmentation is crucial but challenging.
• Convolutional Neural Networks (CNNs) and Transformers excel at local and global feature extraction, respectively.
• Integrating CNNs and Transformers is key for medical image segmentation, but challenges remain in fully utilizing local and global features.

Purpose of the Study:

• To propose a novel CNN-Transformer network, AD2Former, for enhanced medical image segmentation.
• To address limitations in extracting effective local and global features in previous models.
• To improve the segmentation of lesions with unique tissue characteristics and fuzzy boundaries.

Main Methods:

• Developed AD2Former, an encoder-decoder network with an alternating learning encoder and a dual-decoder architecture.
• The alternating learning encoder facilitates real-time interaction between local and global information.
• The dual-decoder architecture employs independent decoding branches with a channel attention module for feature fusion.

Main Results:

• AD2Former demonstrated a strong ability to capture target regions and fuzzy boundaries in medical images.
• Experiments on multi-organ and skin lesion segmentation datasets confirmed the model's effectiveness.
• The proposed designs, alternating learning encoder and dual decoder, significantly contributed to performance gains.

Conclusions:

• AD2Former effectively integrates local and global feature extraction for superior medical image segmentation.
• The novel architecture addresses key challenges in segmenting complex lesion tissues.
• The model shows significant potential for improving diagnostic accuracy in medical imaging applications.