Abstract

Glioblastoma Stage 4 astrocytoma, a highly aggressive and fatal form of brain cancer, demands early detection to improve patient outcomes and extend survival. Accurate brain tumor segmentation from MRI scans plays a crucial role in treatment planning and diagnosis. However, due to the complex nature of MRI, including variations in tumor size, shape, and location, achieving precise segmentation remains a significant challenge in medical image analysis. While deep learning-based models have been widely adopted for brain tumor segmentation, they often face difficulties in capturing directional attributes, which are critical for accurate tumor delineation. Furthermore, the high computational cost associated with complex models can hinder their practical application in clinical settings.
In this research, we propose a novel brain tumor segmentation method that combines multiresolution handcrafted features with convolutional neural network (CNN)-based features to address these challenges. Basically, this framework integrates non-subsampled Shearlet coefficients, which capture directional information at multiple scales, with CNN-extracted features to enhance the segmentation process. The Shearlet transform is particularly effective in representing singularities, such as edges, which are essential for accurate tumor boundary detection. By employing Shannon entropy, the dimensionality of the Shearlet coefficients is selectively reduced by ensuring that only the most relevant features are retained. This feature selection step not only improves segmentation accuracy but also helps reduce the computational load of the model.
To preserve low-level spatial features from the input images, the shortcut connections with identity blocks are incorporated into the architecture. These connections ensure that important features from early layers are not lost during the deeper stages of the network, thus improving the overall segmentation performance. Despite its sophisticated design, the proposed architecture remains relatively simple, with a focus on minimizing computational time, memory usage, and overall costs, making it more feasible for real-time clinical applications.
The performance of the proposed method was evaluated on a dataset with 42 GBM patients, it achieved an average Jaccard score of 0.853, a Dice coefficient of 0.872, and an accuracy of 0.988 in the segmentation task of the whole tumor region. When compared to state-of-the-art segmentation techniques, this approach demonstrates superior performance, especially in capturing tumor details and preserving spatial information. The promising results suggest that the proposed method has significant potential to improve brain tumor segmentation accuracy and, consequently, support better clinical decision-making and treatment outcomes for GBM patients.