S ³Net: a Synthesis-Segmentation-Spiking Network for Alzheimer's disease detection and segmentation

Early and accurate detection of Alzheimer's disease (AD) from Magnetic Resonance Imaging (MRI) scans is crucial for clinical intervention. A novel S ³Net, a Synthesis-Segmentation-Spiking Network, is proposed for this purpose. It integrates synthetic MRI generation, pathology-aware segmentation, and spike-based classification. The Synthesis Network uses a generative adversarial network framework. In this stage, original MRIs are fused with lesion-only patches from disease-relevant regions. This fusion helps preserve high-frequency pathological structures. The generator is trained with adversarial, L2, and Structural Similarity Index Measure (SSIM) losses. These losses ensure that the synthetic images remain realistic and structurally accurate. The Segmentation Network follows an encoder-bottleneck-decoder design with skip connections. It incorporates latent features from both the generator and the discriminator. A hybrid Dice-Binary Cross-Entropy loss is used to enable precise lesion delineation, even in sparsely annotated regions. For classification, a spiking network is employed. It takes fused segmentation and discriminator features and propagates them through Leaky Integrate-and-Fire neurons. This process captures temporal spike dynamics and supports low-power, event-driven computation. On the Open Access Series of Imaging Studies (OASIS) dataset, S ³Net achieves an Accuracy of 95.1%, an F1-score of 93.0%, and an IoU of 82.6%. The proposed S ³Net model outperforms other state-of-the-art methods, demonstrating its effectiveness and clinical viability for automated AD diagnosis.