Texture-Aware StarGAN for CT data harmonisation

TL;DR

CT data harmonization aims to reduce non-biological kernel-induced differences that hinder model generalization. The authors introduce a Texture-Aware StarGAN that performs one-to-many kernel harmonization by conditioning style transfer on target kernels and embedding texture cues via a Multi-Scale Texture Extractor with a self-attention-based aggregation, enforcing $L = L_{baseline} + L_{txt}$. Evaluated on 46,867 chest CT slices from 197 patients across three kernels, the method shows improved radiomic feature alignment and generally favorable deep-feature alignment compared to the baseline StarGAN, though some challenging kernel pairs persist. The work demonstrates that explicit texture information can enhance CT harmonization, offering a promising preprocessing step to improve robustness of downstream radiomics and diagnostic analyses.

Abstract

Computed Tomography (CT) plays a pivotal role in medical diagnosis; however, variability across reconstruction kernels hinders data-driven approaches, such as deep learning models, from achieving reliable and generalized performance. To this end, CT data harmonization has emerged as a promising solution to minimize such non-biological variances by standardizing data across different sources or conditions. In this context, Generative Adversarial Networks (GANs) have proved to be a powerful framework for harmonization, framing it as a style-transfer problem. However, GAN-based approaches still face limitations in capturing complex relationships within the images, which are essential for effective harmonization. In this work, we propose a novel texture-aware StarGAN for CT data harmonization, enabling one-to-many translations across different reconstruction kernels. Although the StarGAN model has been successfully applied in other domains, its potential for CT data harmonization remains unexplored. Furthermore, our approach introduces a multi-scale texture loss function that embeds texture information across different spatial and angular scales into the harmonization process, effectively addressing kernel-induced texture variations. We conducted extensive experimentation on a publicly available dataset, utilizing a total of 48667 chest CT slices from 197 patients distributed over three different reconstruction kernels, demonstrating the superiority of our method over the baseline StarGAN.

Figures (4)

• Figure 1: Our texture-aware StarGAN for CT data harmonization. (a) Our framework includes a generator network $G$ that perform the harmonization, a mapping network $F$ generating style embeddings, a style encoder $E$, and a discriminator network $D$. Texture awareness is enabled by the Multi-Scale Texture Extractor (MSTE) and the aggregation module. (b) MSTE module extracting a textural representation ($\bm{\mathcal{T}}$ or $\bm{\Tilde{\mathcal{T}}}$ from the input image $\bm{x}_{k_i}$ or $\bm{\tilde{x}}_{k_j}$). (c) Dynamic aggregation that combines the extracted representation into a scalar loss function $\mathcal{L}_{txt}$.
• Figure 2: Evolution of FID score during training: (a) $k_1\leftrightarrow{}k_2$, (b) $k_2 \leftrightarrow{}k_3$, (c) $k_1\leftrightarrow{}k_3$, and (d) the average trend across all harmonizations performed by the models. The symbol $\leftrightarrow{}$ denotes bidirectional harmonization.
• Figure 3: Evolution of the percentage of radiomics features with no statistical difference after harmonization: (a) $k_1 \leftrightarrow k_2$, (b) $k_2 \leftrightarrow k_3$, (c) $k_1 \leftrightarrow k_3$, and (d) average trend across all harmonizations performed by the models. The symbol $\leftrightarrow$ denotes bidirectional harmonization.
• Figure 4: Visual comparison of CT images before harmonization and after harmonization using the baseline StarGAN and the proposed texture-aware StarGAN.