Image harmonization and de-harmonization based on singular value decomposition (SVD) in medical domain.

Background: Medical imaging is fundamental to modern clinical diagnostics, providing essential insights for disease detection and treatment planning. However, variations in imaging equipment, protocols, and conditions across institutions lead to inconsistencies in image quality, which hinders diagnostic accuracy and the performance of machine learning models. Although existing harmonization techniques improve image uniformity, they often result in the loss of critical image details. This study presents novel singular value decomposition (SVD)-based harmonization and de-harmonization algorithms, designed to address these challenges by ensuring consistency across diverse imaging conditions, while preserving essential diagnostic information. Methods: The proposed approach utilizes SVD to decompose medical images into multiple frequency bands, allowing for frequency-specific adjustment that enhances both high-frequency details and low-frequency uniformity. The harmonization process begins by splitting red, green, blue (RGB) images into individual channels and applying SVD to extract principal components, enabling the selective enhancement of clinically relevant structures while mitigating variability in brightness and contrast. The de-harmonization method, in contrast, strategically subtracts high-frequency components to remove unwanted noise and preserve significant details. A novel integration of harmonization and de-harmonization processes is employed to optimize image clarity and diagnostic utility. The method's robustness was evaluated through extensive experimentation, including homology (training and testing on the same dataset) and heterology (training on one dataset and testing on a different dataset) experiments. These tests were conducted across multiple datasets-handwritten digit classification (MNIST, USPS), retinal image segmentation [Digital Retinal Images for Vessel Extraction (DRIVE), Choroidal Artery Segmentation Database (CHASE_DB1)], and breast cancer detection (RSNAbreast, INbreast)-with deep learning models employed for performance evaluation. Results: The SVD harmonization and de-harmonization algorithms outperformed traditional methods in image quality and computational efficiency. In homology tests, they achieved 99.21% accuracy on MNIST and 98.7% on USPS. In heterology tests, they scored 98.7% on USPS (trained on MNIST) and 98.46% on MNIST (trained on USPS). For retinal vessel segmentation, AUCs reached 0.976 on DRIVE and 0.982 on CHASE_DB1. For breast cancer detection, AUCs were 0.934 on RSNAbreast and 0.921 on INbreast. Conclusions: The proposed SVD-based harmonization and de-harmonization algorithms present a robust solution to the challenges of image variability in medical imaging. By addressing inconsistencies across different datasets and imaging modalities, while preserving crucial diagnostic information, the techniques enhance the visual quality and clinical utility of medical images. The method's strong performance in both homology and heterology experiments demonstrates its broad applicability and potential to improve the effectiveness of machine learning models in various medical imaging tasks.