Automatic brain MRI tumors segmentation based on deep fusion of weak edge and context features

Abstract

Brain tumors pose a significant health risk to humans. The edge boundaries in brain magnetic resonance imaging (MRI) are often blurred and poorly defined, which can easily result in inaccurate segmentation of lesion areas. To address these challenges, we proposed an Automatic Brain MRI Tumor Segmentation based on deep fusion of Weak Edge and Context features (AS-WEC). First, AS-WEC introduces the Otsu Double Threshold Weak Edges Adaptive Detection (Otsu-WD), which focuses on tumor edge information and differentiates between lesion edges and normal cerebral sulci and gyri. Second, an edge branching network based on the Gated Recurrent Unit (GRU) is constructed to fully preserve the edge context information of the lesion region. Finally, a maximum index fusion mechanism has been designed to incorporate a multilayer feature map, preventing the loss of edge details during the deep feature fusion process. The experimental results demonstrate that the Otsu-WD method outperforms the Canny and TEED algorithms in detecting brain MRI tumor edges. In brain MRI tumor segmentation, AS-WEC delivers a clearer visual segmentation effect compared to the classical UNet++ network and recent models like PVT-Former. On both datasets, AS-WEC demonstrated improvements across multiple metrics. The Dice averaged 92.96%, and the mIoU reached 93.12%, effectively validating the method’s efficacy in brain MRI tumor segmentation. Code and pre-trained models are available at https://github.com/DL-Segment/AS-WEC.git.