Multi-domain medical image translation generation for lung image classification based on generative adversarial...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Lung image classification is crucial for disease diagnosis, with a significant market for AI-assisted tools.
• Existing translation models struggle with attention, key transfer, image quality, and detailed feature generation.
• Generative Adversarial Networks (GANs) offer potential for medical image translation and classification enhancement.

Purpose of the Study:

• To develop an improved medical image translation model addressing limitations in attention and feature generation.
• To enhance lung image classification accuracy using synthetically generated medical images.
• To evaluate the performance of the proposed model in classifying normal, mild, and severe COVID-19 cases.

Main Methods:

• Proposed MI-GAN (Medical Image Generative Adversarial Network) with a key migration branch for multi-domain translation.
• Utilized imbalanced medical image data by selecting key target domain images and establishing a key migration branch.
• Developed a lung image classification model trained on both real and synthetically generated lung CT images.

Main Results:

• MI-GAN successfully translated and generated normal, viral pneumonia, and Mild COVID-19 lung images with improved authenticity and diversity.
• Synthetic image quality metrics (GAN-test: 92.188%, GAN-train: 85.069%) showed significant improvement over other models.
• The classification model achieved 93.85% accuracy for pneumonia diagnosis, with improved sensitivity (96.69%) and specificity (89.70%), and an increased AUC (96.17%).

Conclusions:

• The proposed multi-domain medical image translation model (MI-GAN) effectively enhances key transfer and attention performance.
• MI-GAN successfully synthesizes high-quality medical images, validated on lung CT datasets.
• The synthesized images significantly improve the performance of lung image classification networks for auxiliary diagnosis.