CTMLP: Can MLPs replace CNNs or transformers for COVID-19 diagnosis?

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Computer Vision

Background:

• Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs) dominate COVID-19 diagnosis but have limitations.
• Pure CNNs lack global modeling, while hybrid CNN-ViTs are computationally expensive and parameter-heavy.
• These limitations hinder effective, just-in-time medical diagnosis applications.

Purpose of the Study:

• Introduce CTMLP, a lightweight network for efficient COVID-19 medical image diagnosis.
• Investigate the efficacy of self-supervised learning algorithms for MLP-based models.
• Develop a novel pre-training scheme for medical imaging due to limited large-scale datasets.

Main Methods:

• Proposed CTMLP, a lightweight network combining convolutions and Multi-Layer Perceptrons (MLPs).
• Developed TL-DeCo, a pre-training scheme using transfer learning and self-supervised learning.
• Constructed a guided self-supervised pre-training scheme for efficient lightweight model training.

Main Results:

• CTMLP achieved 97.51% accuracy, 97.43% F1-score, and 98.91% recall without pre-training.
• CTMLP utilizes only 48% of ResNet50 parameters, demonstrating significant efficiency.
• The guided self-supervised scheme improved baseline self-supervised learning by 1-1.27%.

Conclusions:

• CTMLP offers a more efficient alternative to CNNs and Transformers for COVID-19 diagnosis.
• The developed pre-training framework enhances CTMLP's potential for clinical practice.
• The study highlights the viability of lightweight models and guided pre-training in medical AI.