An Evolutionary Attention-Based Network for Medical Image Classification.

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Computer Vision

Background:

• Deep learning models excel in medical image analysis but often lack generalizability across different diseases.
• Current models perform poorly on diseases other than those they were trained on, highlighting a significant challenge.

Purpose of the Study:

• To develop an effective and robust deep learning network for medical image classification with improved generalizability.
• To address the limitations of current models in handling diverse medical conditions.

Main Methods:

• Proposed the densely connected attentional network (DCA-Net) with channel-wise feature map weighting and dense connectivity for efficient information flow.
• Introduced intra-evolution (weight optimization) and inter-evolution (experience sharing between network instances) to enhance model capability and generalizability.
• Developed the evolutionary DCA-Net (EDCA-Net) by integrating these evolutionary strategies.

Main Results:

• EDCA-Net was evaluated on four distinct medical image datasets.
• The proposed EDCA-Net outperformed state-of-the-art methods on three of the four datasets.
• Comparable performance was achieved on the remaining dataset, indicating robust generalizability.

Conclusions:

• EDCA-Net offers a promising solution for generalizable medical image classification.
• The evolutionary approach enhances model robustness and performance across various diseases.
• This network architecture has the potential to advance AI applications in diverse medical diagnostic tasks.