Multiscale Latent Diffusion Model for Enhanced Feature Extraction from Medical Images

TL;DR

The paper addresses the problem of radiomic feature variability caused by CT scanner and protocol differences by introducing LTDiff++, a multiscale latent diffusion framework built on a UNet++ encoder–decoder. The method performs latent-space standardization via a conditional diffusion model, with a three-phase training procedure that enforces $Z_{A'} \approx Z_B$ and preserves structural fidelity through $\mathcal{L} = \|Z_{A'} - Z_B\|^2 + \lambda \|D(Z_{A'}) - B\|^2$, using updates $Z_{t+1} = \sqrt{1 - \beta_t} Z_t + \sqrt{\beta_t} \epsilon$. Evaluation on patient and phantom CT datasets shows higher concordance correlation coefficients (CCC) across radiomic feature groups, indicating improved reproducibility, and a case study demonstrates that synthesized standardized images yield consistent future lung cancer risk predictions. This approach offers a practical pathway to more reliable radiomic analyses and downstream diagnostics in medical imaging.

Abstract

Various imaging modalities are used in patient diagnosis, each offering unique advantages and valuable insights into anatomy and pathology. Computed Tomography (CT) is crucial in diagnostics, providing high-resolution images for precise internal organ visualization. CT's ability to detect subtle tissue variations is vital for diagnosing diseases like lung cancer, enabling early detection and accurate tumor assessment. However, variations in CT scanner models and acquisition protocols introduce significant variability in the extracted radiomic features, even when imaging the same patient. This variability poses considerable challenges for downstream research and clinical analysis, which depend on consistent and reliable feature extraction. Current methods for medical image feature extraction, often based on supervised learning approaches, including GAN-based models, face limitations in generalizing across different imaging environments. In response to these challenges, we propose LTDiff++, a multiscale latent diffusion model designed to enhance feature extraction in medical imaging. The model addresses variability by standardizing non-uniform distributions in the latent space, improving feature consistency. LTDiff++ utilizes a UNet++ encoder-decoder architecture coupled with a conditional Denoising Diffusion Probabilistic Model (DDPM) at the latent bottleneck to achieve robust feature extraction and standardization. Extensive empirical evaluations on both patient and phantom CT datasets demonstrate significant improvements in image standardization, with higher Concordance Correlation Coefficients (CCC) across multiple radiomic feature categories. Through these advancements, LTDiff++ represents a promising solution for overcoming the inherent variability in medical imaging data, offering improved reliability and accuracy in feature extraction processes.

Figures (2)

• Figure 1: Shows how disparities in imaging procedures might result in differences in tumor picture characteristics. The same scanner was used to scan the identical lungman chest phantom. Two distinct image reconstruction techniques were utilized to get CT images kernels appropriately, as the text at the bottom of the pictures shows. Red squares are used to indicate tumors in the images on the left(tumor zoomed in the top left corner). The feature variance in terms of CCC between these two tumors was displayed by the histogram on the right. The potential of extensive radiomic features may be significantly impacted by the detected variations in the tumor pictures.
• Figure 2: Overview of LTDiff++ architechture. $A$ and $Z_A$: non-standard image and its latent vector. $B$ and $Z_B$: standard image and its latent vector. $A'$: standardized image of $A$ that falls in the $B$ domain. $\eta$: Gaussian noise. Given an image pair $(A, B)$ where $A$ and $B$ are non-standard and the corresponding standard images, the model aims to synthesize a new image $A'$ in the domain $B$. The representation learning component leverages a modified UNet++ encoder-decoder structure. This framework is pivotal for learning encoded latent representations of CT images. Concurrently, the target-specific latent-space mapping component is purpose-built for standard image synthesis. It integrates a DDPM model for effective latent space mapping. Here, $Z_A$ represents the latent vector of the non-standard image $A$, $Z_B$ is the latent vector of the standard image $B$, $Z_{A'}$ is the standardized latent vector of image $A$, and $\eta$ denotes Gaussian noise.