Automated MRI Glioma Segmentation Using Deep Learning: A Framework for Sustainable AI-Enabled Clinical Imaging

Niko Hyka; Elda Spahiu; Kita Sallabanda; Dafina Xhako; Albana Shahini; Joan Jani; Rudina Osmanaj; Suela Hoxhaj

Authors

Niko Hyka Department of Diagnostic, University of Medicine, Tirana, Albania https://orcid.org/0000-0002-4645-3844
Elda Spahiu Institute of Applied Nuclear Physics, Tirana, Albania https://orcid.org/0009-0004-5205-8893
Kita Sallabanda Instituto de Radiocirugia, Avanzada, IRCA, Madrid, Spain https://orcid.org/0000-0002-6294-1549
Dafina Xhako Department of Physics Engineering, Polytechnic University of Tirana, Tirana, Albania https://orcid.org/0000-0002-0937-2471
Albana Shahini Department of Diagnostic, University of Medicine, Tirana, Albania https://orcid.org/0009-0005-6179-0485
Joan Jani Department of Physics Engineering, Polytechnic University of Tirana, Tirana, Albania https://orcid.org/0000-0003-0043-9411
Rudina Osmanaj Department of Physics, Faculty of Natural Sciences, University of Tirana, Tirana, Albania https://orcid.org/0000-0002-0367-9560
Suela Hoxhaj Department of Physics Engineering, Polytechnic University of Tirana, Tirana, Albania https://orcid.org/0009-0006-4686-3663

Keywords:

Glioma, Automated Segmentation, MONAI Label, 3D Slicer, Deep Learning, Multimodal MRI, BraTS, Artificial Intelligence, Medical Imaging

Abstract

Artificial intelligence-based brain tumor segmentation is a critical and challenging task in medical image analysis, particularly for gliomas, the most prevalent and aggressive primary brain tumors in adults. Accurate segmentation of glioma sub-regions from multimodal magnetic resonance imaging (MRI) is essential for diagnosis, surgical planning, radiotherapy, and treatment response assessment. However, manual delineation is time-consuming and prone to inter-observer variability due to heterogeneous tumor morphology, indistinct boundaries, complex infiltration patterns, and the three-dimensional nature of MRI data. In this study, we propose an automated deep learning-based framework for multimodal MRI glioma segmentation using MONAI Label, an AI-assisted annotation system integrated within the 3D Slicer platform. The methodology employs convolutional neural network architectures pre-trained on brain tumor datasets and fine-tuned using multimodal MRI inputs, including T1-weighted, contrast-enhanced T1-weighted, T2-weighted, and FLAIR sequences. An active learning strategy is incorporated to iteratively refine segmentation models by combining automated predictions with expert corrections, enabling efficient human-in-the-loop learning. The proposed framework was evaluated on publicly available brain tumor imaging datasets for training and validation. Segmentation performance was assessed using standard evaluation metrics, demonstrating robust and competitive accuracy across tumor sub-regions. Results indicate that integrating deep learning inference with interactive expert feedback significantly reduces manual annotation time while maintaining high segmentation quality. The integration of MONAI Label with 3D Slicer provides a flexible, reproducible, and clinically applicable workflow for automated glioma segmentation. This approach supports efficient dataset annotation and reliable tumor segmentation for both research and clinical applications.