U-Mamba: a hybrid CNN-SSM architecture for accurate segmentation of Alzheimer’s Disease-related brain structures in T1-weighted MRI

Accurate brain segmentation of gray matter structures from T1-weighted MRI is critical for early diagnosis of Alzheimer’s Disease (AD). It also supports reliable monitoring of disease progression. Existing methods face key challenges. They struggle to balance local feature capture, long-range dependency modeling, and computational efficiency. For solving these problems, we propose U-Mamba. It is a hybrid architecture integrating Convolutional Neural Networks (CNNs) and State Space Models (SSMs). Its core goal is the precise segmentation of 122 brain gray matter structures. U-Mamba uses a U-shaped encoder-decoder structure. The encoder fuses CNN layers for local spatial feature extraction. It also incorporates SSM-based blocks Mamba for efficient modeling in long-range dependency. The decoder leverages skip connections. These connections preserve fine-grained spatial information. They also refine segmentation boundaries. We validate U-Mamba on the Alzheimer’s Disease Neuroimaging Initiative also called as ADNI dataset. The model achieves a median Dice Similarity Coefficient (DSC) of 0.9112. It outperforms the previously developed methods, like nnU-Net, SegResNet, and SwinUNETR. Quantitative and qualitative results yield clear evidence. U-Mamba effectively captures subtle structural changes in AD patients. It performs particularly well in atrophy-prone regions. Such regions include the amygdala and entorhinal cortex. The proposed model offers a computationally efficient and accurate tool. It is tailored for clinical AD diagnosis. It also holds potential applications in personalized treatment planning. Additionally, it can support disease progression tracking.