Improving Pre-trained Adult Glioma Segmentation Models using only Post-processing Techniques

Area of Science:

• Neuro-oncology
• Medical imaging analysis
• Artificial intelligence in medicine

Background:

• Gliomas are aggressive adult brain tumors with poor survival rates.
• Accurate multiparametric MRI segmentation is vital for glioma treatment and monitoring.
• Current deep learning models struggle with generalization and introduce segmentation errors.

Purpose of the Study:

• To develop adaptive post-processing techniques for refining glioma segmentation.
• To address limitations of large-scale pre-trained models in brain tumor segmentation.
• To promote computationally fair and sustainable AI strategies in medical imaging.

Main Methods:

• Proposed adaptive post-processing methods to correct segmentation errors.
• Applied techniques to refine segmentations from pre-trained models.
• Evaluated performance on BraTS 2025 segmentation challenge tasks.

Main Results:

• Achieved a 14.9% improvement in the ranking metric for the sub-Saharan Africa challenge.
• Secured a 0.9% improvement for the adult glioma challenge.
• Demonstrated effectiveness in refining segmentations from generalized models.

Conclusions:

• Adaptive post-processing offers an efficient alternative to complex model architectures.
• This approach enhances precision, computational fairness, and sustainability in brain tumor segmentation.
• Shifts focus towards clinically aligned, post-processing strategies for AI in neuro-oncology.