A Bayesian deep segmentation framework for glioblastoma tumor segmentation using follow-up MRIs.

Area of Science:

• Neuroimaging
• Artificial Intelligence in Medicine
• Oncology

Background:

• Glioblastoma (GBM) is an aggressive brain tumor requiring precise measurement for assessment.
• Accurate segmentation of GBM on follow-up magnetic resonance images (MRIs) is critical for evaluating treatment response.
• Existing deep learning models trained on preoperative MRIs show suboptimal performance on follow-up scans due to post-treatment structural changes.

Purpose of the Study:

• To develop and evaluate a novel Bayesian deep segmentation model for accurate GBM sub-region segmentation in follow-up MRIs.
• To address the limitations of current models in handling post-surgical and therapeutic changes in brain structure.
• To enhance clinical decision-making and treatment evaluation for GBM patients.

Main Methods:

• Developed a Bayesian deep segmentation model using 311 follow-up MRIs.
• Integrated Bayesian learning to quantify prediction uncertainty.
• Employed transfer learning to interpret underrepresented textures and spatial details in follow-up MRI data.

Main Results:

• The proposed model achieved high Dice Similarity Coefficient (DSC) scores: 0.833 for FLAIR hyperintensity, 0.901 for enhancing tumor, and 0.931 for non-enhancing necrosis.
• Significantly outperformed existing models in segmenting GBM sub-regions on follow-up MRIs.
• Demonstrated effective identification and correction of misclassifications using uncertainty metrics.

Conclusions:

• The novel Bayesian deep segmentation model accurately segments GBM sub-regions in follow-up MRIs, accounting for structural changes.
• Leveraging uncertainty metrics enhances segmentation accuracy and refines tumor estimates.
• The model shows potential for improving patient management and treatment evaluation in glioblastoma care.