FAENet: Autoencoding-driven Feature Adaptation for Multi-class Brain Tissue Segmentation

The construction of individualized head models from medical imaging data constitutes a critical focus in noninvasive brain stimulation (NIBS) research, with applications encompassing therapeutic electrical stimulation optimization and safety assessment. These computational head models are typically generated by segmenting magnetic resonance imaging (MRI) data into discrete anatomical tissues. However, conventional threedimensional (3D) convolutional neural networks (CNNs), despite their widespread adoption, exhibit high computational costs, parameter redundancy, and overfitting risks. In this study, we propose FAENet, an autoencoder-based architecture that integrates adaptive multi-branch decoding within a 2.5D segmentation framework for multi-class brain tissue segmentation. The network employs parallel decoder branches to enable multiresolution segmentation, with each branch independently optimized according to the textural heterogeneity across distinct anatomical structures to enhance localization accuracy. Experimental results demonstrate that the proposed model achieves a mean Dice similarity coefficient (DSC) of 9 2. 3% with a processing time of 6.63 seconds per volume. This performance indicates accurate reconstruction of tissue boundaries across all segmented classes while maintaining computational efficiency, thereby outperforming conventional 3D approaches.