Harmonizing T1-weighted MRI via disentangled information bottleneck while preserving brain anatomy

1686 

Abstract Submission 

Seonggyu Kim¹, Hanyeol Yang², Yongseon Yoo², Jihwan Min², Jong-Min Lee^1,2,3

¹Department of Electronic Engineering, Hanyang University, Seoul, Korea, Republic of, ²Department of Artificial Intelligence, Hanyang University, Seoul, Korea, Republic of, ³Department of Biomedical Engineering, Hanyang University, Seoul, Korea, Republic of

Seonggyu Kim  
Department of Electronic Engineering, Hanyang University
Seoul, Korea, Republic of

Hanyeol Yang  
Department of Artificial Intelligence, Hanyang University
Seoul, Korea, Republic of

Yongseon Yoo  
Department of Artificial Intelligence, Hanyang University
Seoul, Korea, Republic of

Jihwan Min  
Department of Artificial Intelligence, Hanyang University
Seoul, Korea, Republic of

Jong-Min Lee  
Department of Electronic Engineering, Hanyang University|Department of Artificial Intelligence, Hanyang University|Department of Biomedical Engineering, Hanyang University
Seoul, Korea, Republic of|Seoul, Korea, Republic of|Seoul, Korea, Republic of

Recently, many deep learning models, such as ChatGPT, have been applied to various industrial fields, leveraging their outstanding performance. Despite being an early focus of AI research, the medical field has seen limited utilization of AI models compared to other sectors. One significant factor contributing to this issue is the challenge posed by insufficient data, leading to suboptimal model performance. For instance, transformer models, widely used today, struggle to learn effectively with limited datasets. Additionally, the variation in image value across cohorts, stemming from differences in data collection methods, hinders the robustness of AI in healthcare applications. A clear and effective solution to these challenges is the construction of large-scale datasets. While it may sound cliché, there are generally two approaches to building such datasets. One involves standardizing data collection protocols for all images, which could include standardizing scanners or simply unifying imaging parameters. The other approach is data harmonization through post-processing of collected images. Even after protocol standardization, the need for data harmonization often arises. Harmonization techniques can be broadly categorized into feature level and image level, with methods like ComBat being prevalent at the feature level and generative models at the image level.

Liu et al.[1] employed conditional variational autoencoders (VAE) and adversarial learning for MRI harmonization. They interpreted their model from the perspective of information bottleneck theory, enhancing a traditional information bottleneck-based VAE by introducing assumptions such as the global Markov property of latent variables (beta and theta) given the input and the shared anatomy of T1 weighted and T2 weighted images within the same subject. However, the optimization of such approaches often fails to guarantee maximum compression of latent variables in the information bottleneck, and the assumption that T1 and T2 weighted images share anatomy may be inaccurate for some slices. In this study, we introduce a disentangled information bottleneck [2] loss to ensure maximum compression of the latent variable beta and train a more interpretable harmonization model through simulation.

Our approach, which maps images from the source cohort and target cohort to the same space during harmonization, demonstrates higher structural similarity index (SSIM) scores than traditional methods, indicating improved performance compared to images from the source and target cohorts.

Using this methodology, we successfully validated the creation of a more robust harmonization model, maintaining individual anatomical information while accommodating contrast variations. Nevertheless, there remain several intricate facets regarding the interpretation of latent variable, which may be addressed in the future through the development of explainable encoding methods. Moreover, it is crucial to assess whether these harmonization methods indeed improve the efficacy of diverse downstream tasks.

Exploratory Modeling and Artifact Removal ¹

Other Methods ²

Data analysis

Data Organization

Informatics

Modeling

MRI

Structures

Other - synthesis