This systematic review was conducted in accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) statement. The protocol for this review was not prospectively registered in a systematic review database (e.g., PROSPERO).

Studies were identified through an electronic search of the Web of Science, PubMed, and Scopus databases conducted on 19th January 2026. These databases were selected for their comprehensive coverage of biomedical and scientific literature, ensuring access to a wide array of relevant research at the intersection of neuroimaging, sex differences, and ML.

The search strategy employed keywords and indexed terms including “sex”, “gender”, “dimorphism”, “predict”, “classify”, “disparities”, “difference”, “discriminate”, “distinguish”, “brain”, “magnetic resonance imaging”, “MRI”, “machine learning”, “deep learning”, and “classifier”. These terms were combined using Boolean operators to refine the results. For example, the following string was adapted for each database: (sex OR gender) AND (dimorphism OR predict OR classify OR disparities OR difference OR discriminate OR distinguish) AND (brain) AND (magnetic resonance imaging OR MRI) AND (machine learning OR deep learning).

To ensure transparency and reproducibility, the exact Boolean search strings used for each database are provided below:

To ensure a high level of technical consistency and diagnostic relevance, studies were selected based on the following criteria:

Two reviewers independently screened all titles and abstracts identified through the search. The same reviewers independently assessed full-text articles for eligibility. Disagreements were resolved through consensus, or when necessary, by consultation with a senior author.

To evaluate the internal validity and methodological rigor of the included studies, a structured narrative quality assessment was performed. In the absence of a singular, universally accepted risk-of-bias tool specifically for ML-neuroimaging, we adapted criteria from the PROBAST (Prediction model Risk Of Bias Assessment Tool) and QUADAS-2 frameworks, tailored to address three critical domains:

Across the 35 included studies, the risk of bias was found to be moderate to high. While the transition toward deep learning (DL) has improved the handling of complex data, consistent gaps remain—specifically regarding the reporting of TIV-normalization and the use of external, cross-site validation sets. These gaps suggest that the “ceiling” of sex-classification accuracy may be partially bolstered by methodological artifacts.

The study selection and screening process is illustrated in the PRISMA flow diagram (Figure 1). The initial search across three electronic databases yielded 112 records. Following the removal of 42 duplicate entries, 70 unique records remained for title and abstract screening. Of these, 14 irrelevant studies were excluded, and 56 articles underwent full-text assessment for eligibility. After a detailed review, 21 articles were excluded for not meeting the pre-defined inclusion criteria (e.g., use of 1.5T MRI, focus on clinical populations, or lack of ML-based classification).

Ultimately, 35 articles [4, 9, 16–48] were deemed eligible and included in the final analysis. Data regarding sample population, age, morphometric features, ML algorithms, and sex-classification accuracy were extracted from these studies and are summarized in Table S1. Our quality assessment revealed a critical methodological landscape: while the field has transitioned toward powerful DL architectures, there remain significant gaps in inference validity. Notably, only a minority of high-accuracy studies explicitly corrected for TIV or employed external, cross-scanner validation sets. This suggests that the reported “performance ceiling” of sex classification may be partially bolstered by allometric scaling and site-specific artifacts.

This systematic review investigated the efficacy of various ML algorithms in classifying sex based on gray matter (GM) morphometric features. Overall, the classification models exhibited high accuracy in distinguishing between sexes, ranging from 56.0% to 98.06%. The highest performing method (98.06%) employed a specialized DL approach termed Multi-Layer 3D Convolution Extreme Learning Machine (MCN-ELM) [25], while the lowest accuracy (56.0%) was observed using a support vector machine (SVM) [31].

Diffusion-weighted imaging (DWI) examines water diffusion patterns to assess white matter microstructural integrity and brain connectivity. Five of the included studies utilized diffusion-based features for sex classification [19, 23, 43, 46, 47].

Fractional anisotropy (FA) and DL: Xin et al. [43] (2019) achieved a peak accuracy of 93.3% using FA maps paired with a 3D CNN; notably, their comparative SVM model reached only 78.2% on the same data. Chen et al. [19] (2024) reported a similar accuracy of 91.0% using FA with a 2D CNN, which unexpectedly outperformed both 3D CNN and Vision Transformer (ViT) architectures. This indicates that DL excels at extracting subtle connectivity signatures that traditional ML (TML) misses.

fMRI captures dynamic brain activity by measuring blood-oxygen-level-dependent (BOLD) signals. This review evaluated both task-based fMRI, which maps activity during specific cognitive challenges, and resting-state fMRI (rs-fMRI), which assesses intrinsic functional connectivity.

Task-based fMRI: Two studies utilized task-based paradigms, reporting moderate classification performance. Leming and Suckling (2021) achieved an AUC of 0.783 using an emotion-processing task, while Xu et al. [44] (2020) reported an accuracy of 75.86% using a language-based paradigm.

A key finding across the literature is the significant improvement in classification accuracy achieved through the integration of multiple MRI sequences. Five studies demonstrated that fusing morphological and functional data—including T1-weighted, T2-weighted, diffusion-weighted, and functional (resting-state and task-evoked) MRI—consistently outperformed single-modality models.

Incremental gains: The earliest evidence for this trend came from Wang et al. [39] (2012), who observed that combining GM volume with rs-fMRI increased accuracy to 89.1%, surpassing the 86.1% achieved using GM volume alone.

A comparative analysis of the included studies reveals distinct patterns between TML and DL architectures. While TML models—including SVM, RFs, and gradient boosted trees—demonstrated high performance, DL models consistently established the upper bounds of classification accuracy.

TML: Among studies utilizing TML, the peak accuracy reached was 96.77%, achieved by Luo et al. [28] (2019) through their HSRC approach. SVM was the most prevalent algorithm, utilized in 17 studies. Zhang et al. [42] (2020) reported the highest SVM-based accuracy (96.6%) by integrating structural and functional features. However, TML performance varied significantly, with some SVM models reporting accuracies as low as 55.0% [31, 46], highlighting a high sensitivity to feature selection and preprocessing.

In head-to-head evaluations, the choice of algorithm often dictated performance. Sanchis-Segura et al. [34] (2022) compared five algorithms [linear discriminant analysis (LDA), LR, MARS, RF, and SVM] in a sex- and age-matched cohort (n = 876), finding that SVM outperformed the others with an accuracy of 90.0% when using uncorrected GM volume.

However, the assumption that DL always outperforms TML was challenged by Ebel et al. [9] (2023). In their study using the Study of Health in Pomerania (SHIP) dataset, LR (95.78%) outperformed a CNN [9]. This was attributed to the use of z-score normalization on GM voxels, suggesting that rigorous feature engineering in TML can occasionally match or exceed the automated feature extraction of DL models. This dichotomy underscores that while DL generally offers higher ceilings, TML remains highly effective when paired with optimized preprocessing.

Given that brain maturation and aging involve distinct neurobiological processes, this review assessed how ML accuracy varies across the lifespan.

Children and adolescents (under 23 years): Eight studies focused on younger populations [17, 18, 30, 32, 35, 37, 38, 46]. The highest accuracy reported was 97.3% in a large-scale study of approximately 11,000 children (aged 9–12 years) from the Adolescent Brain Cognitive Development (ABCD) study [17]. This suggests that sex-specific neuroanatomical patterns are already highly distinguishable by late childhood.

To reach further stratification power, the included studies were characterized by developmental stage and morphometric constraints. A clear performance gradient emerged when stratifying by age: studies focusing on the adolescent stratum (9–12 years) and young adult stratum (20–35 years) reported the highest mean accuracies (97.3% and 98.06%, respectively). In contrast, studies spanning the full adult lifespan (14–90 years) showed increased variance and a slightly lower performance ceiling (95.78%), likely due to age-related cortical atrophy. Furthermore, when characterizing groups by TIV, studies that did not stratify for head size reported consistently higher accuracies. This suggests that the “power” of sex classification is significantly mediated by the homogeneity of the age and size strata being analyzed.

The most consequential finding regarding the validity of these models is the role of confounding factors, primarily TIV. Because male brains are, on average, 8–12% larger than female brains [10], models trained on uncorrected volumetric data may achieve “near-perfect” accuracies by exploiting global scaling cues rather than intrinsic multivariate morphology. Our review found that reporting on TIV correction was alarmingly limited [9, 34]. As Sanchis-Segura et al. [34] demonstrated, sex differences appear “large” when uncorrected but become significantly attenuated once TIV is controlled. If a model “cheats” by using head size as a proxy for sex, it is not identifying neurobiological sex differences, but rather allometric scaling. This distinction is vital: if the predictive signal is size-driven, the field is not yet in a position to use these models as evidence for stable, mechanistically informative dimorphism.

The transition toward DL has pushed classification boundaries toward a 98% ceiling [25]. Architectures such as 3D CNNs excel by automatically learning complex, non-linear spatiotemporal patterns. However, DL models showed higher sensitivity to methodological artifacts. While TML, specifically SVMs, remains the most utilized due to transparency [18, 23], DL is the primary driver of state-of-the-art precision. Yet, without robust explainability, it remains unclear if these “black box” models are identifying biological truth or “predictive noise”.

Integrating multiple MRI sequences consistently outperformed single-modality models [40]. Fusing structural (T1w, T1-weighted image), diffusion (dMRI), and functional (fMRI) data allows for a holistic view of sex-based variation. However, the increased accuracy of multimodal models (up to 96.6%) may also amplify the aggregation of confounding signals—such as motion artifacts in fMRI paired with size-bias in T1w data—unless rigorous cross-modal confound control is applied.

A critical methodological tension exists between traditional statistical-driven and ML-driven applications. While statistical approaches prioritize the identification of significant regional differences (inference), ML prioritizes the maximization of classification accuracy (prediction). Our analysis highlights that these two approaches do not always yield the same prioritization of brain features [9, 34]. For instance, an ML model might prioritize a combination of subtle, non-significant features—such as cortical folding patterns or functional connectivity weights—to reach high accuracy. This discrepancy suggests that ML may “prioritize” features that are statistically weak on their own but computationally powerful in aggregate, potentially exploiting “predictive noise” that lacks true inferential validity.

High accuracy does not equate to biological generalizability. A recurring limitation is the absence of cross-site or cross-scanner validation. Models trained on single-site datasets (e.g., HCP, UKB) may inadvertently learn “site-specific signatures”—such as scanner gradients or head-coil configurations—rather than universal biological signals. Furthermore, the reliance on WEIRD (Western, Educated, Industrialized, Rich, Democratic) populations introduces demographic biases. An accuracy of 98% may reflect overfitting to site-specific noise and demographic homogeneity, failing when applied to more diverse groups.

Crucially, these findings do not imply deterministic brain dimorphism. High classification accuracy is a testament to the power of multivariate statistics to find patterns in overlapping data, not an endorsement of binary biological essentialism. Individual brains are more accurately described as a “mosaic” [6] of features. Transparent reporting and cautious interpretation are essential to prevent the misuse of high-accuracy ML results in contexts that demand binary certainty, where nature provides a spectrum of variability.

The high classification accuracies identified in this review—reaching up to 98.06%—necessitate a move beyond viewing sex-classification as a purely technical challenge. Our findings reveal specific methodological trends that carry significant ethical weight.

Our review found that T1-weighted imaging and DL architectures (e.g., MCN-ELM) produce the highest accuracies. However, these results must be interpreted with caution. There is a profound ethical risk that these “near-perfect” statistical separations are used to support essentialist claims—the idea that male and female brains are “opposite” or “dimorphic”. As noted in our discussion of the “brain mosaic”, high accuracy in a multivariate model does not negate the extensive overlap in individual brain features. Misrepresenting these statistical patterns as deterministic labels could be misused to justify gender-based discrimination in education, employment, or forensic contexts.

Our findings show a heavy reliance on datasets like the HCP, UKB, and ABCD. Because these cohorts predominantly sample WEIRD populations, there is a direct ethical risk of creating inequitable algorithms. If a model achieves 98% accuracy on a high-income, Western cohort but is then deployed in a global health context without cross-cultural validation, it may produce biased or invalid results for underrepresented groups. The “robustness” reported in current literature is, therefore, demographically “fragile”, potentially reinforcing systemic healthcare disparities.

A critical ethical concern arising from our analysis is the lack of TIV correction in several high-accuracy studies. When models “cheat” by using global head size as a proxy for sex, they are not identifying neurobiological sex differences; they are identifying physical scale. Attributing this to “brain sex” is ethically problematic, as it presents a technical confound as an innate biological truth, potentially leading to flawed theories about cognitive or behavioral differences.

To mitigate these risks, we recommend that future neuroimaging-ML research adopt the following mandatory reporting standards:

1.

Mandatory TIV disclosure: Researchers must explicitly report whether and how they corrected for TIV to ensure accuracy reflects morphology, not just size.

A primary limitation of this systematic review is the methodological heterogeneity regarding confound management, particularly TIV. Since male brains are, on average, 8–12% larger than female brains, models that do not rigorously normalize for TIV—a gap identified in several high-accuracy studies—may be capturing allometric scaling rather than intrinsic neuroanatomical dimorphism. This “size-confounding” significantly inflates accuracy and obscures true biological signals.

Furthermore, the field’s heavy reliance on a few large, shared public datasets (e.g., HCP, UKB) introduces a profound population bias. These datasets represent WEIRD cohorts, which lack the diversity necessary to claim universal biological validity. The absence of cross-site validation in many studies suggests that models may be overfitting to site-specific noise—such as scanner manufacturer gradients or specific preprocessing idiosyncrasies—rather than generalizable neurobiology. This “site-signature” effect means that a 98% accuracy on one dataset might drop to chance levels when applied to another [44].

The absence of prospective protocol registration in a repository such as PROSPERO is a notable limitation. Although the review followed a rigorous, predefined internal protocol verified by multiple independent authors to minimize selective reporting bias, the lack of public registration may impact the transparency of the study selection process. Additionally, while our search was expanded to include Web of Science, PubMed, and Scopus to capture the interdisciplinary intersection of neuroscience and engineering, the exclusion of other technical databases (e.g., IEEE Xplore) remains a potential constraint on the absolute comprehensiveness of the literature surveyed. Finally, the restriction to English-language publications may exclude culturally diverse findings that could further challenge the “binary” narrative of brain sex. These combined artifacts suggest that current literature may provide an inflated sense of certainty regarding the categorical distinction of the human brain.

This systematic review reveals that while ML can achieve peak accuracies of 98.06% in sex classification, these results must be viewed with significant skepticism. The high performance is likely bolstered by methodological artifacts, including inadequate TIV correction and the learning of dataset-specific scanner signatures. Rather than confirming a deterministic binary, these models identify population-level statistical trends that are heavily influenced by global head size and demographic homogeneity.

We conclude that the human brain is better described as a mosaic of overlapping features rather than two distinct categories. For ML to move from a “pattern detector” to a valid neuroscientific tool, future research must prioritize mandatory TIV disclosure, cross-site validation, and the use of demographically diverse datasets. Without these rigorous standards, high classification accuracy remains a statistical achievement of limited biological and clinical utility.