The challenge the team set out to address is a familiar one in neuroimaging. Brain MRI data are extraordinarily high-dimensional, and many clinically meaningful signals, whether they mark disease progression or a phenotypic contrast, are subtle enough to escape the human eye. Traditional statistical approaches such as voxel-based morphometry analyze the brain one voxel at a time and rely on linear modeling assumptions, which limits their ability to capture nonlinear patterns or coordinated effects spanning multiple regions. Standard machine learning methods like support vector machines and random forests can capture nonlinearity, but they demand pre-defined feature engineering, forcing researchers to guess in advance which brain features matter. Deep learning, and convolutional neural networks in particular, removes that bottleneck by learning features directly from raw image data, from local edge and texture patterns up to higher-level structural configurations.

Yet most deep learning pipelines for spatial localization operate on entire three-dimensional volumes, which is computationally expensive and typically requires very large training cohorts. SliceMap takes a different route. Instead of modeling the whole brain at once, it extracts slices from the three orthogonal anatomical planes, sagittal, coronal, and axial, and trains an independent two-dimensional convolutional neural network on each slice. Binary classification performance, measured with the Matthews correlation coefficient, becomes the yardstick: slices that classify the two groups well are presumed to carry the strongest discriminative information. The best-performing slice from each plane is then subjected to occlusion-based attribution analysis, in which small patches of the image are systematically masked and the resulting change in the model’s prediction is measured, revealing exactly which pixels drive the decision.

The network architecture itself is deliberately compact. It comprises three convolutional blocks, each pairing a three-by-three convolutional layer with batch normalization, a rectified linear unit activation, and two-by-two max pooling, with the number of feature maps increasing from eight to sixteen to thirty-two across the blocks. A fully connected layer with 128 hidden units, layer normalization, and dropout precede the final classification output. Each model was trained in PyTorch on a single NVIDIA A30 GPU using the AdamW optimizer, and hyperparameters such as weight decay and learning-rate scheduling were tuned on a sagittal mid-slice model before being fixed for all other slice-specific models. The sagittal mid-slice model reached a mean Matthews correlation coefficient of 0.546 across ten independent runs, a solid baseline for a task as subtle as sex classification.

What sets the study apart is the rigor of its validation. In a controlled simulation, the researchers artificially injected a signal into a known cubic region of the brain, modifying voxel intensities in female scans within that cube while leaving everything else untouched. When the pipeline was rerun, the Matthews correlation coefficient spiked to a perfect 1.00 with zero variance, but only for the slices intersecting the injected region; adjacent slices showed no change at all. This proved the pipeline can detect and localize a signal at a known location. A permutation experiment provided the complementary check: when sex labels were randomly shuffled, model performance collapsed to near zero across all planes, confirming that the networks were learning genuine biological structure rather than spurious correlations.

With the framework validated, the team turned to the real data. Slice-wise performance profiles showed minimal discriminative power at the edges of the brain and steadily increasing performance toward central slices, where tissue coverage is greatest. The single best-performing slices achieved Matthews correlation coefficients of roughly 0.60 in each plane, with accuracies near 80 percent. Occlusion heatmaps generated on those slices revealed a concentrated high-attribution zone in the upper-central portion of the axial view, and crucially, the best sagittal slice passed directly through that same region while the corresponding coronal slice showed elevated attribution at the same coordinates. Converging evidence from all three viewing angles allowed the team to define a candidate three-dimensional region of 24 by 24 by 24 voxels.

That candidate region corresponds anatomically to the anterior cingulate cortex and adjacent medial structures, including portions of the corpus callosum genu. To confirm the finding, the researchers trained a small three-dimensional convolutional network exclusively on patches extracted from the candidate region and compared it against an identically sized control region drawn from a low-attribution area. The candidate region achieved a Matthews correlation coefficient of 0.457 and an accuracy of 72.9 percent, substantially outperforming the control region’s 0.304 and 65.0 percent. The result aligns with a body of prior neuroimaging literature: voxel-based morphometry studies have repeatedly reported proportionally greater anterior cingulate gray matter volume in women, a large comparative study found the anterior cingulate cortex significantly larger in females in both humans and mice, and midsagittal corpus callosum area has been reported as significantly larger in females after controlling for brain size.

The team also explored whether combining information across viewing planes improves classification. A multi-slice model fed the three best-performing slices from the sagittal, coronal, and axial planes achieved a Matthews correlation coefficient of 0.719 and an accuracy of 85.9 percent, a clear gain over any single-plane model. This suggests the selected slices capture complementary discriminative information distributed across anatomical views. A separate analysis comparing slice-level performance metrics with occlusion and GradCAM attribution scores found moderate to strong correlations in most planes, indicating that the computationally cheap Matthews-based slice selection serves as a reasonable surrogate for more expensive attribution-based localization, though the two criteria capture partially distinct aspects of the underlying signal.

The researchers emphasize that SliceMap is designed to be task-agnostic. Its core components, slice-wise network training, performance-guided slice selection, and occlusion-based attribution, make no assumptions specific to sex classification and could in principle be applied to any binary contrast in neuroimaging, from disease versus healthy comparisons to other phenotypic groupings. A preliminary test on distinguishing Alzheimer’s disease patients from cognitively normal subjects produced modest and unstable results, but the authors attribute this to the limited sample size of only 371 matched pairs after confounder matching, rather than to a fundamental limitation of the pipeline itself.

Limitations remain. Localization analyses were not conditioned on prediction confidence, the optimal slice combination across planes has not yet been systematically optimized, and a head-to-head comparison with fully three-dimensional architectures is still an open question. The authors suggest that confidence-aware localization, joint multi-plane slice optimization, and aggregation of larger multi-cohort datasets are natural next steps. Still, the demonstration that a fleet of small two-dimensional networks can triangulate a biologically plausible three-dimensional brain region, and that the finding replicates across independent validation checks, positions SliceMap as a practical and efficient new instrument for the ongoing effort to translate the overwhelming dimensionality of brain imaging into interpretable anatomical insight.

Subject of Research: A 2D slice-based deep learning pipeline for detecting discriminative brain regions in MRI

Article Title: SliceMap: a binary classification-driven 2D pipeline for detecting discriminative candidate regions in brain MRI

Article References: Jiang, X., Wu, Z., & Qin, Z. S. (2026). SliceMap: a binary classification-driven 2D pipeline for detecting discriminative candidate regions in brain MRI. Frontiers in Neuroimaging, 5, Article 1819355. https://doi.org/10.3389/fnimg.2026.1819355

Image Credits: AI Generated

DOI: 10.3389/fnimg.2026.1819355

Keywords: brain MRI, deep learning, convolutional neural networks, occlusion attribution, sex differences, anterior cingulate cortex, corpus callosum, neuroimaging, ADNI, binary classification, spatial localization, SliceMap