Overview
Brain tumor segmentation is a challenging problem in medical image analysis. The objective is to generate precise masks that accurately identify brain tumor regions in fMRI screenings. In this paper, we propose a novel Attention Gate (AG) model for brain tumor segmentation, which combines an edge detecting unit with an attention gated network to highlight and segment salient regions from fMRI images. This integration eliminates the need for explicit external tissue localization and traditional classification methods used in classical computer vision techniques. AGs can be seamlessly incorporated into deep convolutional neural networks (CNNs), requiring minimal computational overhead while significantly enhancing sensitivity scores. Our results demonstrate that the combined edge detector and attention gated mechanism is a robust method for brain segmentation, achieving an Intersection Over Union (IOU) score of 0.81.
Services
The concept of using edge information to enhance processing is well-established. From early methods to more recent models like the Inf-Net module, these approaches have effectively provided constraints that guide feature extraction for segmentation tasks. To learn the edge representation, we feed low-level features with moderate resolution into the proposed Edge Attention (EA) module. This module uses a 1x1 convolutional filter to produce a filtered mapping of the original image. The gated module is trained using standard Binary Cross Entropy.
We propose exploring the benefits of Attention Gates (AGs) in medical imaging, specifically for image segmentation, by introducing a grid-based gating mechanism. This approach allows attention coefficients to be more precisely targeted to local regions, enhancing segmentation accuracy.
Results
Our results demonstrate that combining the edge detector with an attention gated mechanism provides an effective method for brain tumor segmentation, achieving an Intersection Over Union (IOU) score of 0.81. We trained the model using weighted Focal Loss Binary Cross Entropy to address class imbalance and variations in tumor representation within the database.
To further evaluate the performance, we reported mean IOU, precision, recall, and F1-Score. Confusion matrices for the subject-wise 10-fold cross-validation approach on the testing data from the augmented dataset are presented in Figure 5. The IOU for the testing data is 81%.
Link to the code: https://github.com/Timothy102/Brain-FMRI
Link to the paper: https://arxiv.org/abs/2107.03323