Multibranch CNN With MLP-Mixer-Based Feature Exploration for High-Performance Disease Diagnosis

Area of Science:

• Medical imaging analysis
• Artificial intelligence in healthcare
• Deep learning for diagnostics

Background:

• Conventional convolutional neural networks (CNNs) have limitations in feature exploration due to restricted receptive fields and biased extraction.
• This compromises the performance of deep neural networks (DNNs) in medical diagnosis.
• Optimal DNN design is crucial for high-performance healthcare diagnostics.

Purpose of the Study:

• To propose a novel feature exploration network, the manifold embedded multilayer perceptron (MLP) mixer (ME-Mixer), for enhanced disease diagnosis.
• To leverage both supervised and unsupervised features for improved diagnostic accuracy.
• To develop a generalizable network plugin compatible with existing CNN architectures.

Main Methods:

• A manifold embedding network is utilized to extract class-discriminative features.
• Two MLP-Mixer-based feature projectors are employed for global feature encoding.
• The ME-Mixer is designed as a plugin module for integration into existing CNNs.

Main Results:

• Comprehensive evaluations on two medical datasets demonstrated significant improvements in classification accuracy.
• The proposed ME-Mixer enhanced performance compared to various DNN configurations.
• The approach achieved these enhancements with acceptable computational complexity.

Conclusions:

• The ME-Mixer effectively addresses the feature exploration limitations of traditional CNNs.
• The proposed network offers a significant advancement in deep learning-based medical diagnosis.
• ME-Mixer provides a versatile and efficient solution for improving diagnostic performance in healthcare.