Importance of conditioning in latent diffusion models for image generation and super-resolution

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Image Processing

Background:

• Photon-counting computed tomography (CT) offers advanced imaging capabilities.
• High-resolution CT data is crucial for accurate diagnosis but can be limited by noise and acquisition constraints.
• Synthesizing and enhancing medical images is vital for dataset augmentation and improving image quality.

Purpose of the Study:

• To investigate the application of latent diffusion models (LDMs) for synthesizing and enhancing photon-counting chest CT images.
• To evaluate LDMs for dataset augmentation and super-resolution (SR) tasks.
• To assess the potential of LDMs in supporting diagnostic accuracy and accessibility to high-resolution imaging data.

Main Methods:

• A framework combining a variational autoencoder (AutoencoderKL) and a denoising diffusion model was developed.
• Experiments explored various conditioning strategies, including segmentation masks and class labels (lung, soft tissue).
• Multiple loss functions and conditioning approaches were tested across generative and SR tasks.

Main Results:

• Unconditioned LDMs generated anatomically inaccurate images lacking clinical interpretability.
• Conditioning with segmentation masks and anatomical labels significantly improved structural fidelity.
• The best image generation achieved MS-SSIM of 0.7135 and PSNR of 24.53; SR tasks reached MS-SSIM of 0.85 and PSNR of 27.31.

Conclusions:

• Latent diffusion models demonstrate significant potential for augmenting and enhancing photon-counting chest CT images.
• Conditioning LDMs with anatomical information preserves structural integrity and minimizes hallucinated anatomy.
• These models offer a pathway for controllable, high-fidelity image synthesis in clinically relevant applications.