ED-Net: Edge-Enhanced Frontal Lobe Segmentation Network Based on Multi-Scale Feature Extraction

Cognitive impairment is a common and serious consequence of stroke, often associated with damage to the frontal lobe, which is responsible for essential cognitive functions. However, accurately segmenting the frontal lobe in MRI images remains challenging due to blurred anatomical boundaries, complex lesion morphology, and significant inter-subject variability. To address these challenges, we propose an edgeenhanced frontal lobe segmentation network (ED-Net) based on multi-scale feature extraction. ED-Net incorporates a Multi-scale Edge Feature Extraction (MEF) module that captures boundary information at various scales using dilated convolutions, and an Edge Preservation (EP) module that suppresses noise while retaining fine edge details through gated convolutional blocks. Furthermore, an Attention-guided Multi-scale Convolution Module (AGM) is embedded within the encoder to enhance spatial contextual understanding. Experiments conducted on a brain perfusion MRI dataset from Liaoning Provincial People's Hospital demonstrate that ED-Net achieves an average Dice coefficient of $\mathbf{8 3. 1 5 \%}$, confirming its effectiveness in frontal lobe segmentation. The proposed method improves segmentation accuracy for complex frontal lobe structures, offering robust technical support for the clinical assessment of post-stroke cognitive impairment.