Synthesizing Chest X-Ray Pathology for Training Deep Convolutional Neural Networks

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Machine Learning

Background:

• Medical datasets frequently exhibit class imbalance, with common conditions overrepresented and rare diseases underrepresented.
• Privacy concerns hinder the aggregation of large-scale medical datasets across healthcare institutions.

Purpose of the Study:

• To address data imbalance and privacy issues in medical imaging by synthesizing pathology.
• To evaluate the effectiveness of synthesized medical images in improving diagnostic model performance.

Main Methods:

• A deep convolutional generative adversarial network (DCGAN) was implemented to synthesize chest X-ray images.
• A modest-sized labeled dataset was used as the basis for image synthesis.
• Deep convolutional neural networks (DCNNs) were trained using a combination of real and synthesized images for pathology classification across five classes.

Main Results:

• DCNNs trained with both real and synthesized images demonstrated superior performance compared to those trained solely on real images.
• The improved performance was attributed to dataset balancing achieved through DCGAN-synthesized images, particularly augmenting underrepresented rare disease classes.
• Synthesized data effectively addressed the challenge of limited examples for certain pathologies.

Conclusions:

• Synthesizing medical images using DCGANs is a viable strategy to overcome data imbalance and privacy challenges in medical AI.
• The integration of synthesized images enhances the accuracy of DCNNs in classifying pathologies, offering a promising approach for rare disease detection.
• This method holds potential for improving the robustness and generalizability of diagnostic AI models in healthcare.