MACSDE: Multi-Head Autocorrelated Convolutional Squeeze Densely-Connected Encoder for Alzheimer’s Volumetric Biomarker Prediction

Alzheimer’s Disease (AD) is a progressive neurodegenerative disorder that manifests with subtle structural changes in key brain regions, often occurring years before clinical symptoms appear. Early and accurate identification of volumetric biomarkers, such as the hippocampus, entorhinal cortex, and ventricles, from MRI scans is essential for timely diagnosis and monitoring disease progression. However, existing deep learning models primarily focus on classification tasks and often fail to capture the region-specific anatomical nuances necessary for precise volumetric estimation. To address this limitation, we propose a novel hybrid architecture,MACSDE (Multi-head Autocorrelated Convolutional Squeeze Densely-connected Encoder), which combines spatially-aware multi-head correlated convolution with dense connectivity and attention-based recalibration. This architecture is designed to capture region-specific features, enhance inter-regional interactions, and prioritize Alzheimer-relevant channels for accurate volumetric predictions. The model incorporates a brain-wide feature extraction layer, a cognitive integration module, and a regression block tailored for volumetric output. Extensive experiments conducted on benchmark MRI datasets demonstrate the superior performance of MACSDE. Our results show that MACSDE achieves a 15% reduction in MAE, a 20% improvement in RMSE, and significantly higher correlation scores compared to existing methods. Specifically, the model outperforms state-of-the-art approaches like MRNN, PLSTM-Z, and Deep RNN, which show MAE values of 0.143, 0.305, and 0.059, respectively, while MACSDE reduces this to 0.0508. Similarly, RMSE improves from 3.951 (MRNN) and 2.929 (PLSTM-Z) to 1.125 for MACSDE, showing enhanced consistency in biomarker volume prediction. MACSDE offers competitive, state-of-the-art volumetric AD biomarker prediction performance in the 2024-2025 literature, as demonstrated by comparisons with recent advanced architectures.