Diagnose Harmonisation (DHARM): An open-source tool for the harmonisation and analysis of batch effects in MRI-derived measures.
This paper introduces DHARM, an open-source Python toolbox that provides a standardized, multi-criteria framework for diagnosing, applying, and evaluating MRI harmonization methods to effectively reduce technical batch effects while preserving biologically meaningful signals.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer
Magnetic resonance imaging, or MRI, has revolutionized how we see inside the human body, offering a non-invasive window into the brain's structure and function. For decades, researchers relied on single machines in single locations to gather data, ensuring that every scan was taken under identical conditions. However, as the scientific community has sought to answer bigger questions about aging, disease, and development, the need to combine data from many different hospitals and scanners has become essential. The challenge is that every MRI machine is slightly different, and every hospital follows its own specific procedures for taking pictures. These technical differences act like a persistent background noise, creating systematic variations in the data that have nothing to do with the biology of the person being scanned. If a researcher tries to compare brain measurements from a patient in London with one from Oxford, the differences might reflect the machines rather than the patients, potentially leading to false conclusions.
To solve this, scientists have developed statistical methods to smooth out these technical discrepancies, a process known as harmonisation. The goal is to remove the noise introduced by the equipment while keeping the true biological signal intact. Yet, for a long time, there was no standard way to check if these methods were actually working. Different research groups used different tests to judge success, making it difficult to know which method was best for a specific situation or whether a method had accidentally erased important biological details while cleaning up the noise. This lack of a common language for evaluation left researchers guessing about the quality of their combined datasets.
In response to this uncertainty, a team of researchers from Oxford and Nottingham has introduced a new open-source software tool called DHARM, which stands for Diagnose Harmonisation. Rather than offering a single magic fix, this tool provides a comprehensive toolkit for diagnosing exactly what kind of technical noise exists in a dataset and then rigorously testing whether a chosen method has successfully removed it. The researchers built DHARM to act as a universal referee, allowing scientists to apply various harmonisation techniques and then immediately see, through a series of clear visual reports and statistical checks, how well each method performed. The tool is designed to be flexible, capable of handling data from a single snapshot in time as well as data collected from the same people over multiple years, ensuring that the complex relationships between repeated measurements are respected.
The team demonstrated the power of DHARM by applying it to real-world data from several large studies, including the UK Biobank and a specialized dataset where healthy volunteers were scanned on multiple different machines. They took raw, unharmonised data and ran it through five different harmonisation approaches, ranging from simple statistical corrections to more complex models that account for the shape of the data. The results revealed that while all the methods were good at reducing the most obvious differences in average measurements, they did not all perform equally well on other fronts. Some methods successfully aligned the average values but left behind differences in how much the data varied from person to person, while others struggled to preserve the natural relationships between different brain features.
Crucially, the tool showed that a method appearing successful on one measure could fail on another. For instance, a technique might make the average brain volume look the same across different scanners, but it could inadvertently distort the natural variation in those volumes or obscure the link between brain size and age. By using a battery of different tests, DHARM allowed the researchers to see these trade-offs clearly. They found that methods which adjusted both the average values and the spread of the data tended to perform better overall, but even the best methods left some residual differences depending on the specific characteristics of the scanners involved. The tool also highlighted that in studies where the same people were scanned multiple times, it is vital to account for the fact that repeated scans from one person are not independent; ignoring this can lead to misleading conclusions about how consistent the measurements are.
The demonstration confirmed that there is no single "best" method for every situation. Instead, the right choice depends on the specific data at hand and the biological question being asked. A method that works well for correcting simple shifts in average values might not be suitable for data where the differences are more complex or where the biological signal is non-linear. DHARM empowers researchers to move beyond a one-size-fits-all approach by providing a transparent, evidence-based way to compare methods side-by-side. It allows them to ask not just whether the noise has been reduced, but whether the true biological story remains clear and intact.
By integrating these diverse checks into a single, user-friendly workflow, the researchers have provided the field with a way to make harmonisation a more reliable and reproducible part of the scientific process. The tool does not just tell scientists that their data is cleaner; it shows them exactly how it has changed and what has been preserved. This level of detail is essential for building confidence in large-scale studies that combine data from dozens of sites, ensuring that the conclusions drawn about human health are based on biology rather than the quirks of a particular machine. The introduction of DHARM marks a shift toward a more rigorous standard in neuroimaging, where the success of data cleaning is measured by a complete picture of the data's health, rather than a single, potentially misleading number.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.