Deep Learning based Hippocampal segmentation in MRI for early detection of Alzheimer’s disease

Accurate segmentation of the hippocampus from structural MRI plays a pivotal role in the early diagnosis and follow-up of Alzheimer’s disease, as even subtle changes in hippocampal structure are among the earliest markers of neurodegeneration. In recent years, deep learning models particularly encoder–decoder architectures have driven significant progress in medical image segmentation. However, their effectiveness depends heavily on architectural choices such as decoding strategy, feature refinement, and the capability to highlight clinically relevant regions. This study presents a systematic comparison of four segmentation models U-Net, Attention U-Net, Basic SegNet, and a newly developed Symmetric SegNet implemented within a unified training and evaluation framework. All models were optimized using a hybrid Binary Cross-Entropy and Dice loss, enabling fair assessment across diverse network designs. The experimental findings show that the incorporation of attention mechanisms improves focus on the hippocampal region, leading to sharper boundary detection, whereas a symmetric encoder–decoder pathway enhances structural continuity in regions with low tissue contrast. Quantitative evaluation using Dice coefficient and Intersection over Union (IoU) further confirms that architectural symmetry and targeted attention significantly improve segmentation fidelity. Overall, the comparative results highlight the strengths and limitations of each architecture and demonstrate that carefully designed symmetry and attention-guided refinement can substantially enhance hippocampal segmentation. These insights contribute to the development of more reliable neuroimaging pipelines, supporting early detection and monitoring of Alzheimer’s disease in clinical and research settings.