Efficient 4D fMRI analysis via spatio-temporal screening and region-aware feature extraction for template-free brain disorder classification

<i>Objective.</i>Functional magnetic resonance imaging (fMRI) is crucial for identifying neurological disorder biomarkers, but current deep learning methods face some limitations. Template-dependent methods reliant on fixed brain atlases lack inter-subject specificity and generalizability due to fixed anatomical priors. Emerging template-free models, which process raw data directly, often separate spatial and temporal processing. This approach discards temporal continuity, which encompasses key characteristics such as the smooth and correlated nature of neural dynamics over time. To address these limitations, we propose a novel axial slice-centric model that jointly models spatiotemporal representations through end-to-end processing of native 4D fMRI data. This eliminates template dependency while preserving intrinsic brain activity patterns.<i>Approach.</i>Our framework redefines 4D fMRI analysis by decomposing it into 3D spatiotemporal manifolds along the axial axis, enabling joint learning of spatial and temporal features and preserving individualized structure organization. A hierarchical encoder extracts local spatiotemporal interactions within each slice, progressively aggregating information to capture multi-granularity neural patterns. To maintain temporal continuity and computational efficiency, a differentiable TopK operation adaptively selects informative slices and time points, balancing computational demands with long-range temporal dependencies.<i>Main results.</i>Experimental results on the ADNI dataset (324 subjects, for classifying early mild cognitive impairment and normal controls )) and a private disorder of consciousness dataset (164 subjects) demonstrate the effectiveness of our 4D fMRI framework in classifying both neurodegenerative and consciousness disorders. Specifically, on the ADNI dataset, our proposed model achieves 97% classification accuracy with over 25% reduction in floating-point operations compared to baseline methods. On the private dataset, our model outperforms state-of-the-art approaches by 5% accuracy. Visualization of slice-level attention maps identify biomarkers consistent with previous research, demonstrating that our template-free framework can discover biomarkers comparable to those identified by template-dependent methods.<i>Significance.</i>Our joint spatiotemporal modeling framework, enabled by axial slice-centric decomposition of 4D fMRI data while preserving temporal continuity, achieves excellent complexity-accuracy trade-offs for brain disorder analysis. Biomarker visualization confirms its template-free capability to identify clinically-relevant neural patterns, offering an efficient and interpretable solution for 4D fMRI-based diagnosis.