Advancing Glioma Segmentation: A Robust 3D Residual Attention U-Net Framework for Multimodal MRI Images

Brain tumors are abnormal growths of cells within the brain, posing significant health challenges.Among the different types of brain tumors, glioma, which originates from supportive glial tissue, is notably concerning due to its low survival rate compared to others.The prognosis of glioma hinges on various factors, including its location, size, and degree of extension.Therefore, accurate segmentation of glioma is crucial for both diagnosis and treatment planning.Manual segmentation by radiologists or specialized clinicians is prone to variability and is time-intensive.To automate tumor segmentation, a '3D Residual Attention U-Net' architecture is presented in this study.By integrating spatial and channel attention mechanisms, the model enhanced the feature representation.The "Vanishing Gradient" phenomenon in conventional 'U-Net' was solved using a series of residual directed blocks.Moreover, the use of a modified loss function called 'Focal-Dice' helped alleviate challenges associated with noisy labels and class imbalance in pixel segmentation tasks.The proposed architecture achieves a Dice coefficient of 0.9002 and an Intersection over Union (IoU) metric of 0.8272, demonstrating its efficacy in segmenting brain tumors.These results are obtained within a short training time, showcasing the model's efficiency.Overall, this research contributes to streamlining medical image analysis processes, ultimately enhancing patient care outcomes in glioma diagnosis and treatment.