Cerebral Edema Image Segmentation Based on Improved U-Net Modeling

In this study, for the clinical needs of accurate segmentation of brain edema region, an improved U-net model based on deep learning is proposed, which achieves a significant improvement in the performance of medical image segmentation by integrating three core modules, namely, multi-scale perception, lightweight convolution and attention mechanism. First, the Min-ASPP module is introduced in the encoder part, which solves the feature dilution problem of traditional ASPP in tiny lesion recognition by effectively capturing multi-scale contextual information through an improved void space pyramid structure; second, the decoding path is reconstructed by using the GSConv module, whose hybrid architecture of depth-separable convolution and standard convolution maintains the feature expression capability while keeping the model parameter number is reduced; finally, the AG attention mechanism is embedded at the jump junction, which strengthens the feature differentiation between edematous regions and normal tissues through adaptive feature calibration. After the histogram equalization and elastic deformation enhancement in the data preprocessing stage, the model shows excellent performance indexes on the test set: the Dice coefficient reaches 94.32%, and the IoU value reaches 91.93%. The clinical application value of this achievement lies in the fact that the high-precision and high-efficiency automatic segmentation algorithm can assist clinicians to quickly locate the scope of edema, quantitatively assess the progress of the disease, and provide a reliable basis for surgical planning and efficacy assessment, which is of great practical significance in promoting the intelligent diagnosis and treatment of stroke, craniocerebral injury and other diseases.