Introduction

Deep learning methods in neuroimaging have largely focused on population-level classification and group-based inference, limiting their applicability for individualised clinical interpretation. Convolutional neural networks (CNNs) are well suited to modelling complex spatial patterns in structural MRI; however, their use for single-subject analysis remains limited. This study introduces a technically novel normative CNN framework designed to enable personalised, single-subject interpretation of T1-weighted structural MRI without reliance on diagnostic group comparisons or case–control statistics.

Methods

A three-dimensional convolutional autoencoder was implemented to learn normative neuroanatomical representations from healthy adult T1-weighted MRI data. Approximately 100 scans were selected from the publicly available IXI dataset to construct the normative reference model. Images underwent standard pre-processing including skull stripping, spatial normalisation, and intensity standardisation. The trained CNN was subsequently applied to individual subjects to generate voxel-wise reconstruction error maps, representing deviations from learned normative structure. Quantitative regional deviation scores were computed by aggregating voxel-level errors within anatomically defined brain regions, enabling subject-specific neuroanatomical profiling.

Results

The proposed framework produces continuous, voxel-wise deviation measures across the entire brain volume from a single T1-weighted MRI scan. Reconstruction error maps provide spatially localised representations of neuroanatomical variability relative to normative structure learned from the IXI reference cohort. Regional aggregation yields quantitative subject-specific deviation profiles, supporting interpretable individual-level assessment. Preliminary analyses demonstrate stable model training and consistent deviation patterns across healthy reference scans, indicating the technical feasibility of CNN-based normative modelling for single-subject inference.

Conclusions

This work presents a single-subject CNN framework for personalised structural MRI analysis. By shifting deep learning applications from group-level classification to normative individual-level deviation mapping, the proposed approach supports personalised medicine and precision neuroimaging paradigms. The framework provides a foundation for future large-scale validation, longitudinal modelling, and clinical translation of personalised neuroimaging methodologies.