A Generative Model of Lung CT Conditioned on Radiomics Features

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Pathology

Background:

• Traditional deep learning image generation models lack control over specific output properties.
• Controlling image texture features is crucial for applications like medical data synthesis.

Purpose of the Study:

• To develop a deep learning model capable of generating images with user-specified texture features.
• To utilize a diffusion transformer architecture conditioned on texture features for controlled image synthesis.

Main Methods:

• A diffusion transformer architecture was employed for image generation.
• Texture features (autocorrelation, inverse difference from Gray-Level Co-Occurrence Matrix) were used as conditional inputs.
• The model was trained and evaluated on lung patches from a public CT database.

Main Results:

• The model successfully generated lung parenchyma-like image patches.
• Generated images showed strong agreement and low variability with conditional texture features (autocorrelation, inverse difference).
• High concordance correlation coefficients (0.9962 for autocorrelation, 0.9402 for inverse difference) were achieved.

Conclusions:

• The diffusion transformer model effectively generates images with controlled texture features.
• This approach supports highly controlled data generation for diverse applications, particularly in medical imaging.
• The model's ability to align generated image textures with conditional inputs is validated.