Multi-branch CNN and grouping cascade attention for medical image classification

Area of Science:

• Medical Image Analysis
• Computer Vision
• Artificial Intelligence

Background:

• Visual Transformers (ViT) show promise in medical imaging but struggle with small datasets and high computational demands.
• Existing ViT models often have redundant computations within multi-head self-attention (MHSA).

Purpose of the Study:

• To develop an efficient hybrid network, Eff-CTNet, for medical image classification.
• To address the limitations of ViT in terms of performance on small datasets and computational efficiency.

Main Methods:

• Proposed an efficient hybrid network (Eff-CTNet) by alternating Convolutional Neural Networks (CNNs) and Transformers.
• Introduced a Group Cascade Attention (GCA) module to reduce redundancy and improve attention diversity in Transformers.
• Developed an Efficient CNN (EC) module to enhance local feature extraction.

Main Results:

• Eff-CTNet demonstrated advanced classification performance on three public medical image datasets.
• The proposed model achieved superior results with significantly lower computational cost compared to existing methods.
• The GCA and EC modules effectively improved attention diversity and local detail extraction.

Conclusions:

• Eff-CTNet offers a computationally efficient and high-performing solution for medical image classification.
• The hybrid approach effectively leverages the strengths of both CNNs and Transformers.
• This network is well-suited for practical clinical applications due to its efficiency and accuracy.