Latent Diffusion Models with Image-Derived Annotations for Enhanced AI-Assisted Cancer Diagnosis in Histopathology

Area of Science:

• Digital Pathology
• Medical Imaging
• Artificial Intelligence

Background:

• Supervised AI methods in histopathology require extensive annotated datasets, which are often scarce.
• Synthetic data generation offers a promising solution to augment limited real-world datasets.
• Latent diffusion models can produce high-quality synthetic images but typically need detailed text prompts, unavailable in histopathology.

Purpose of the Study:

• To develop a method for generating synthetic histopathology images using AI.
• To address the challenge of limited annotated data for training AI diagnostic models.
• To evaluate the effectiveness of AI-generated synthetic data in improving AI model performance.

Main Methods:

• Proposed a novel method to construct structured textual prompts from automatically extracted image features for latent diffusion models.
• Experimented with the PCam dataset, containing tissue patches annotated as healthy or cancerous.
• Evaluated synthetic image quality using Fréchet Inception Distance (FID) and assessed pathologist's ability to distinguish real from synthetic images.

Main Results:

• Including image-derived features in prompts improved FID by 88.6 compared to using only diagnostic labels.
• Pathologists demonstrated limited ability to differentiate synthetic from real histopathology images (median sensitivity/specificity of 0.55/0.55).
• Synthetic data generated using the proposed method effectively trained AI models for diagnostic tasks.

Conclusions:

• AI-driven synthetic data generation, guided by image features, can overcome annotation limitations in histopathology.
• The developed method enhances the quality and utility of synthetic data for AI training in cancer diagnostics.
• Synthetic histopathology images show potential for robustly training AI models, supporting diagnostic applications.