Multi-objective Bayesian optimization with enhanced features for adaptively improved glioblastoma partitioning and...

Area of Science:

• Neuro-oncology
• Medical Imaging
• Machine Learning

Background:

• Glioblastoma (GBM) is an aggressive brain tumor with complex microstructures and vascular patterns.
• Delineating GBM subregions is crucial for developing targeted therapies.
• Current unsupervised learning methods struggle with reliability and clinical relevance in GBM subregion identification.

Purpose of the Study:

• To develop a robust framework for reproducible and clinically meaningful glioblastoma subregion delineation.
• To establish the clinical relevance of identified subregions through survival prediction.
• To optimize hyper-parameters efficiently for both clustering stability and outcome significance.

Main Methods:

• Introduction of a multi-objective Bayesian optimization (MOBO) framework.
• Integration of a Feature-enhanced Auto-Encoder (FAE) with customized stability and clinical significance losses.
• Modeling of reproducibility and clinical relevance using distinct Gaussian Processes (GPs) within the MOBO framework.

Main Results:

• The MOBO framework efficiently optimizes hyper-parameters for FAE architecture and clustering.
• Achieved a balance between clustering stability and clinical relevance using bespoke losses.
• Demonstrated robust MRI subregion delineations and statistically validated survival predictions.

Conclusions:

• The proposed MOBO framework offers an efficient and effective solution for glioblastoma subregion delineation.
• This approach enhances the reliability and clinical applicability of radiomic feature analysis in GBM.
• The method accelerates hyper-parameter tuning, leading to improved diagnostic and prognostic capabilities.