Deep learning models can spot patterns in medical images that are hard for people to see. In one open-access study, researchers used diffusion MRI (an MRI method that tracks how water moves in brain tissue) to test whether brain scans contain consistent signals linked to sex assigned at birth. The models performed well, and key signals came from central white matter tracts (the brain wiring that helps regions communicate). At the same time, the work does not tell us what causes these differences, and it does not test any clinical diagnostic use.
Brain sex differences detected in white matter microstructure
The study analyzed diffusion MRI from 1031 healthy young adults (471 male and 560 female, ages 22 to 37) in the Human Connectome Project dataset. The goal was not to argue that one brain is better. It was to test whether there are detectable, repeatable patterns in tissue microstructure that differ by sex when brains are compared in a standardized way.
To do this, the team trained three different deep learning architectures (2D CNN, 3D CNN, and a vision transformer) on maps derived from diffusion MRI. Across models and metrics, the reported test AUC ranged from 0.92 to 0.98 (AUC is a standard way to summarize how well a classifier separates two groups across thresholds). The authors interpret this as evidence of consistent sex-related differences in white matter microstructure in this cohort, as described in Scientific Reports: high sex classification performance across three models.
What diffusion MRI adds beyond brain size and shape
Many past debates about sex differences in the brain focus on macroscopic features such as volume or thickness. Diffusion MRI is different. It measures how water diffuses through tissue, which can indirectly reflect properties of microstructure in white matter. In the paper, the authors used three common diffusion-derived metrics (fractional anisotropy, mean diffusivity, and mean kurtosis) and registered the maps to a standard template to reduce the influence of overall brain size and contour.
The study also used “occlusion analysis” to estimate which white matter regions contributed most to the classification signal. The details vary across models and metrics, but the authors report consistent signals in central white matter, and they discuss the corpus callosum as one region that repeatedly mattered in their analyses.
For a plain-language overview, NYU Langone summarizes the work and notes that the analysis aimed to reduce the influence of overall brain size and shape. In the paper, this is done by registering diffusion maps to a standard template to reduce macroscopic size and contour effects: analysis aimed to reduce reliance on overall brain size and shape.
Why sex-based brain patterns matter for research quality
A key point is not whether a model can label a brain scan, but what this implies for research practice. The Scientific Reports paper notes that prevalence differs by sex for several neurological and neuropsychiatric conditions, including autism spectrum disorder and multiple sclerosis. If studies treat one sex as the default, they risk missing patterns that could matter for understanding disease mechanisms or treatment response.
This aligns with broader research standards. NIH guidance emphasizes that “sex as a biological variable” should be considered in study design, analysis, and reporting for vertebrate animal and human research. In other words, sex can be a source of systematic variation that affects whether findings generalize well. See the NIH Office of Research on Women’s Health summary here: sex as a biological variable should be factored into research designs.
If you want related context on how brain imaging can reveal subtle patterns, this is conceptually similar to other work on brain signals and task structure, such as this Gromeus article on how the brain adapts activity patterns during complex cognitive tasks.
What this study cannot tell you about individuals
This work is about group-level patterns in a specific dataset, not a personal brain “test.” Even strong classification performance does not tell us what causes the differences, whether they are driven by biology, environment, or both. It also does not imply that any particular cognitive trait is explained by these imaging patterns.
NYU Langone also highlights an important limitation: while the AI models could report that patterns differ, they could not say which sex was more likely to have which features. In the dataset used, sex labels were self-reported, and the authors report no cases where self-reported sex differed from genetic sex.
If you want a different angle on what can help brain health in everyday life, this is separate from sex differences research, but it is often more actionable. For example, a Penn State study suggests that everyday movement can support cognitive health.
Limitations and quality of evidence
The core evidence here is strong for the specific claim “these models can distinguish sex in this cohort using diffusion MRI microstructure.” But it is still early for broader claims. The sample is limited to young adults in one major dataset, and the work does not connect these patterns to clinical outcomes. Replication in other age groups and datasets, and careful work separating sex-related biology from gender-related lived experience, will matter for interpretation.
Sources and related information
Scientific Reports – Deep learning with diffusion MRI as in vivo microscope reveals sex-related differences in human white matter microstructure – 2024
This peer-reviewed article is the primary source for the claims that deep learning can classify biological sex from diffusion MRI with high test AUC in a Human Connectome Project cohort, and that key signals appear in white matter microstructure after template registration reduces macroscopic differences.
Scientific Reports – Author Correction (citation fix) – 2024
Scientific Reports published an Author Correction on 05 August 2024 that removes an incorrect NODDI citation from the discussion. It does not change the main results summarized in this article.
NYU Langone – Artificial intelligence tool detects sex-related differences in brain structure – 2024
NYU Langone’s press release provides a plain-language description of the same study, including the claim that the analysis aimed to reduce reliance on overall brain size and shape. It links to the Scientific Reports paper and describes key limitations about what the models can and cannot infer: analysis aimed to reduce reliance on overall brain size and shape.
NIH ORWH – NIH policy on sex as a biological variable – n.d.
The NIH ORWH page explains that NIH expects SABV to be factored into research designs, analyses, and reporting, supporting the broader point that sex can be an important variable for study rigor and generalizability.
