GAN-based medical image small region forgery detection via a two-stage cascade framework

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Generative adversarial networks (GANs) are used for medical image enhancement.
• GAN-based attacks, like CT-GAN, can subtly tamper with medical images (e.g., CT scans) by injecting or removing lesions.
• Detecting these small, subtle tampered regions is challenging for existing methods.

Purpose of the Study:

• To propose a robust two-stage cascade framework for detecting small region forgeries in GAN-manipulated medical images.
• To accurately classify tampered CT images and precisely locate the forged regions.

Main Methods:

• A local detection stage using sub-images with depthwise separable convolution, residual networks, and attention mechanisms to identify tampered areas.
• A global classification stage employing Gray-Level Co-occurrence Matrix (GLCM) and infinite-dimensional hyperplanes for classification.
• Combining local detection results into a heatmap for global analysis.

Main Results:

• The proposed method demonstrates superior performance compared to state-of-the-art detection techniques.
• The framework successfully detects and localizes subtle GAN-based forgeries in CT scans.
• Experimental results validate the effectiveness of the two-stage approach.

Conclusions:

• The developed two-stage cascade framework offers a highly effective solution for detecting GAN-based medical image forgeries.
• The method addresses the challenge of identifying small, difficult-to-detect tampered regions in medical scans.
• This work contributes to enhancing the integrity and reliability of medical imaging analysis.