Magnetic Resonance Imaging (MRI) is crucial for medical diagnostics, providing insights into anatomical structures and diseases. However, variability in acquisition protocols across sites leads to inconsistencies in image appearance, caused by differences in scanner hardware, imaging parameters and patient demographics. This variability hinders the effectiveness of deep learning models for tasks like segmentation and classification.
To tackle these challenges, MRI harmonisation techniques aim to standardise image distributions for better comparability. While traditional methods like histogram matching and supervised domain adaptation have been used, they often require paired datasets and aren't scalable. A recent study in Medical Image Analysis presents an advanced unsupervised harmonisation framework using normalising flows, allowing MRI standardisation without needing source domain images during inference.
Addressing Domain Shifts in MRI Data
MRI acquisition protocols vary significantly between different imaging sites and devices, leading to discrepancies in image intensity, contrast and resolution. These inconsistencies introduce a domain shift, which can adversely affect deep learning models trained on a specific dataset. As a result, models optimised for one imaging site often perform poorly when applied to data acquired elsewhere, limiting their applicability in multi-centre studies. Moreover, traditional harmonisation methods often rely on impractical assumptions, such as access to paired source and target images or predefined target domains.
Conventional harmonisation techniques aim to reduce domain shift by mapping images from different distributions onto a shared reference. Histogram matching, for example, aligns intensity distributions across datasets, while statistical methods such as Combat attempt to model and correct for scanner-induced variability. Although these methods offer some improvements, they often fail to capture complex distributional differences and struggle to generalise across diverse imaging conditions. Machine learning-based approaches, including generative adversarial networks (GANs) and style transfer models, provide more flexible harmonisation strategies. However, they typically require training on both source and target data, limiting their feasibility in real-world scenarios.
The proposed normalising flow-based framework seeks to overcome these limitations by learning the statistical characteristics of a given source domain and applying them to target images without requiring direct access to the source data at test time. This source-free, unsupervised harmonisation method enables deep learning models to perform consistently across different imaging sites, improving their generalisability and clinical applicability.
Normalising Flows for Source-Free Harmonisation
Normalising flows are a class of generative models that transform complex probability distributions into simpler, well-characterised distributions through a series of invertible transformations. In the context of MRI harmonisation, a normalising flow model is trained to learn the probability distribution of a specific source domain. This enables the generation of images that match the statistical properties of the source dataset, facilitating a standardised representation of MRI scans from different sites.
Must Read: DL Harmonisation in Abdominal CT Imaging: Advancing Radiomics Reliability
The proposed framework consists of three key steps. First, a normalising flow model is trained to capture the statistical properties of the source domain. This process involves transforming the complex image distribution into a latent space where it follows a simple, known distribution. Second, a harmoniser network is trained to reconstruct images from augmented versions of the source domain, providing an initial mapping function. This network learns to correct for variations in image contrast, brightness and intensity while preserving structural integrity. Finally, during inference, the harmoniser is refined using the pre-trained normalising flow model to ensure that the transformed images conform to the learned source distribution. This test-time adaptation process allows for flexible harmonisation without requiring retraining for each new target domain.
Unlike traditional methods that require knowledge of the target domain in advance, this approach is entirely source-free and task-agnostic. By eliminating the need for explicit domain supervision, it offers a scalable solution for harmonising multi-site MRI datasets. Additionally, because the harmoniser network operates in an unsupervised manner, it can be applied across different tasks, including segmentation, classification and regression.
Evaluation and Generalisation Across MRI Tasks
The effectiveness of this harmonisation strategy has been evaluated across multiple MRI applications, including adult and neonatal brain segmentation, as well as neonatal brain age estimation. In segmentation tasks, images harmonised using normalising flows demonstrated improved consistency across different imaging sites, leading to superior model performance compared to existing harmonisation techniques. The approach was particularly effective in mitigating domain shifts in multi-centre studies, where variations in imaging protocols typically degrade deep learning model performance.
Furthermore, the proposed method was validated on both T1-weighted and T2-weighted MRI datasets, demonstrating its adaptability across different imaging modalities. Traditional harmonisation approaches often struggle to maintain structural details when applied to different MRI sequences, but normalising flow-based harmonisation ensures that essential anatomical features are preserved. This is crucial for applications such as brain segmentation, where fine-grained structural details play a significant role in accurate medical analysis.
Another key advantage of this framework is its generalisability across different populations. In neonatal brain age estimation, for example, the method improved performance by aligning the distribution of images from different sites, ensuring that models trained on one dataset could be effectively applied to unseen data. This capability is particularly beneficial for large-scale clinical studies, where patient demographics and imaging conditions may vary significantly.
MRI harmonisation is crucial for the reliability of deep learning models across diverse imaging datasets. The proposed normalising flow-based framework offers a source-free, task-agnostic approach that functions entirely unsupervised. By learning the statistical distribution of a source domain and adapting images, this method ensures consistent medical image analysis across various sites and conditions.
This approach is advantageous as it eliminates the need for paired training data or prior knowledge of the target domain, making it ideal for handling unseen data distributions. It enhances the practical applicability of deep learning in medical imaging, particularly in multi-centre studies with diverse patient populations. As MRI technology evolves, such advancements in harmonisation will be vital for improving diagnostic accuracy and expanding AI-driven healthcare solutions.
Source: Medical Image Analysis
Image Credit: iStock
