Automatic Skull Stripping Using Multidimensional Multi-input Multi-output U-Net Model for Alzheimer’s Disease

Abstract

Skull stripping is a fundamental step in analyzing magnetic resonance imaging (MRI) scans, which play a crucial role in disease diagnosis such as Alzheimer’s disease (AD). Alzheimer’s is a progressive neurological disorder with no known cure. Early and precise diagnosis of AD is essential for timely intervention to help slow its progression. Although manual brain segmentation from MRI is accurate, it requires expert knowledge, experience, and time investment. Therefore, many automated brain segmentation algorithms have been introduced so far. The U-Net model has recently gained significant attention due to its exceptional volumetric medical image segmentation performance. This study presents a novel multidimensional multi-input multi-output U-Net (MIMO-U-Net) model for more efficient brain extraction. The model is multidimensional because it works with both 2D and 3D datasets. This architecture uses a dropout regularization technique with varying dropout rates across different layers. The concatenation connections are also used to combine high-level features with up-sampled output. The dropout regularization and concatenation help in enhancing the model performance. A refined loss function is also proposed by combining Dice loss and categorical focal loss. The MIMO-U-Net is trained and tested using a T1-weighted ADNI brain MRI dataset. The results indicate that MIMO-U-Net surpasses most existing techniques by offering better accuracy and notable quantitative and qualitative outcomes. In addition, the MIMO-U-Net showcases substantial computational efficiency during execution. Evaluation metrics, comprising the Dice coefficient, specificity, and sensitivity, corroborate the model’s performance with precise scores of 0.992, 0.999, and 0.995, respectively.