Prototype-Driven Class-Conditional Synthesis for High-Quality Chest X-ray Image Generation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Medical image generation faces challenges with data scarcity and imbalanced class distributions, leading to overfitting and poor quality.
• Existing methods struggle to adequately model tail classes in long-tailed datasets, compromising synthetic image realism and diversity.

Purpose of the Study:

• To propose a novel Class Prototype-Driven Diffusion Model (CPDM) for improved class-conditional image synthesis on long-tailed chest X-ray datasets.
• To address the limitations of current image generation techniques in handling severe class imbalance and data scarcity in medical imaging.

Main Methods:

• Introduced a Class Prototype Bank to store representative feature vectors for each class, aiding in the extraction of limited tail class features.
• Integrated cross-attention mechanisms between image features and class-specific prototypes within a diffusion model framework.
• Developed CPDM to capture fine-grained class features, enhancing realism and diversity in generated chest X-ray images.

Main Results:

• CPDM achieved state-of-the-art performance with the lowest Fréchet Inception Distance (FID) of 31.600 and highest Inception Score (IS) of 2.842.
• Demonstrated effectiveness in mitigating class imbalance and data scarcity issues in chest X-ray imaging.
• Downstream classifier experiments showed a 17.22% improvement in mean Area Under the Curve (mAUC) for 14 thoracic diseases when trained on a mixed dataset with only 1% real images.

Conclusions:

• CPDM effectively enhances class-conditional image synthesis for long-tailed medical datasets, particularly chest X-rays.
• The proposed method successfully addresses data scarcity and class imbalance, leading to more realistic and diverse generated images.
• CPDM shows significant potential for improving diagnostic accuracy in medical imaging tasks through data augmentation.