Progressive attention integration-based multi-scale efficient network for medical imaging analysis with application...

Area of Science:

• Medical Imaging Analysis
• Deep Learning
• Artificial Intelligence in Healthcare

Background:

• Medical imaging data often suffers from imperfections, leading to insufficient feature learning.
• Accurate analysis of medical images is crucial for disease diagnosis, such as in COVID-19 detection.

Purpose of the Study:

• To develop a novel deep learning framework, the multi-scale efficient network (MEN), for enhanced medical imaging analysis.
• To address challenges of insufficient feature learning in imperfect imaging data.
• To improve the accuracy and generalization ability of COVID-19 diagnostic models.

Main Methods:

• The proposed multi-scale efficient network (MEN) integrates various attention mechanisms for progressive feature extraction.
• A fused-attention block with squeeze-excitation (SE) attention is used for fine-grained detail extraction.
• A multi-scale low information loss (MSLIL)-attention block with efficient channel attention (ECA) is employed to enhance semantic correlations and compensate for global information loss.

Main Results:

• The MEN framework was evaluated on two COVID-19 diagnostic tasks.
• The proposed method achieved competitive and high accuracies of 98.68% and 98.85% in COVID-19 recognition.
• The MEN model demonstrated satisfactory generalization ability compared to other advanced deep learning models.

Conclusions:

• The developed multi-scale efficient network (MEN) offers a robust solution for medical imaging analysis, particularly for COVID-19 diagnosis.
• The integration of attention mechanisms effectively extracts detailed and semantic features, overcoming data imperfections.
• The MEN framework shows significant potential for accurate and reliable medical image diagnostic applications.