Image harmonization: a review of deep learning methods and evaluation metrics for effective harmonization

The integration of different modalities of data affects the model generalizability and also the data consistency across multiple dataset sources. These variations arise due to the differences in imaging protocols and environmental conditions, that can be resolved by facilitating image harmonization techniques. The major goal of this review paper is to explore deep learning (DL)-based approaches for image harmonization and to guide researchers in identifying suitable frameworks. The paper begins by providing a comprehensive overview of image harmonization strategies. A literature search, utilizing PRISMA guidelines, is conducted to select the most relevant papers from 2019 to 2025. The underlying DL architectures include deep image harmonization, Generative Adversarial Networks (GANs), Variational Autoencoders (VAEs), and attention mechanisms. Furthermore, the performance metrics and dataset evaluation are done for further understanding of data performance on different DL architectures. Finally, this review highlights the major benefits and the limitations of harmonization techniques with future perspectives.