A novel detection-guided approach for multimodal brain tumor segmentation

Glioblastoma is one of the most aggressive brain tumors due to its fast growth and penetration into brain tissue. Glioblastoma is diagnosed and treated with MRI, but its diverse form provides distinct complications. Advanced segmentation approaches are needed due to the tumor's intricacy, poor prognosis, and high recurrence rate. This work aims to build a deep learning-based approach to segment the tumor's three regions—Whole Tumor (WT), Tumor Core (TC), and Enhanced Tumor (ET)—for Computer Assisted Diagnosis. Innovative detection-guided segmentation uses a cutting-edge detection model to identify the probable entire tumor region and a hybrid CNN-transformer network to segment the clipped discriminative region. Integrating the transformer with CNN in the encoder design improves the baseline U-net architecture in this study. The transformer encodes tokenized picture patches from CNN feature maps to capture long-range contextual information, while the decoder integrates upsampled features for low-level details. Feature fusion was used to fuse decoder outputs to improve model learning from multiple hierarchical levels. Finally, a unique fusion method combines local and global segmentation. Our findings demonstrate improved glioblastoma segmentation from MRI scans, with dice scores of 89.09, 85.85, and 91.56 for the WT, TC, and ET regions. Our theories, representations, and strategies for brain tumor segmentation are significant. Addressing medical diagnostics' problems and complications by automating this segmentation process helps clinicians.