Optimization of glioma segmentation using 3D U-Net++ in MRI surgical planning and patient safety outcomes

Abstract

The main goal of this study is to develop and evaluate a novel 3D U-Net++ convolutional neural network for accurate segmentation of glioma sub regions in MRI scans, aiming to enhance surgical planning, targeted radiotherapy, and patient safety. Precise segmentation of glioma sub-regions is a persistent challenge in neuro-oncology due to substantial morphological variability across patients. To address this, we introduce an automatic segmentation model based on a 3D U-Net++ architecture with dense skip connections, which improves spatial feature extraction and the delineation of tumor boundaries. Utilizing volumetric data from the BraTS 2020 benchmark, the model automatically segments three clinically relevant substructures: tumor core, the enhancing tumor, and whole tumor. The integration of dense connections with 3D convolutional layers facilitates the detection of subtle tissue variations, including necrosis and edema. Quantitative evaluation demonstrates that the proposed 3D U-Net++ surpasses conventional architectures such as standard U-Net and DeepMedic in Dice coefficient, sensitivity, and specificity, yielding more homogeneous and continuous segmentations while reducing manual and semi-automatic annotation efforts. This approach supports advanced clinical decision making and workflow automation, and offers potential for application to other tumor types or integration into real-time clinical practice.