VMAM-NET: A Model Agnostic Meta-Learning Network for Rare De Novo Glioblastoma Diagnosis

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Oncology

Background:

• Glioblastoma diagnosis is challenging due to rarity and data scarcity.
• Intratumoral heterogeneity complicates accurate identification.
• Advanced deep learning methods are needed for improved diagnostic accuracy.

Purpose of the Study:

• To introduce VMAM-NET, a hybrid deep meta-learning model for enhanced glioblastoma diagnosis.
• To address data-scarce settings in rare brain tumor identification.
• To improve the efficiency and accuracy of neuroimaging analysis for aggressive brain tumors.

Main Methods:

• Developed VMAM-NET, a hybrid model combining VGG-16 feature extraction with model-agnostic meta-learning (MAML).
• Utilized VGG-16 pre-trained on Astrocytoma data for domain-specific feature acquisition.
• Applied MAML for rapid adaptation to few-shot glioblastoma learning tasks on four MRI datasets.

Main Results:

• VMAM-NET achieved high performance with 98.69% training accuracy and 96.71% testing accuracy.
• The model obtained an F1-score of 0.9694, outperforming traditional deep learning and meta-learning approaches.
• Gradient-based Class Activation Maps (Grad-CAM) provided interpretability by highlighting tumor regions.

Conclusions:

• VMAM-NET offers a scalable and clinically feasible solution for glioblastoma diagnosis in data-limited environments.
• The model demonstrates the potential of data-efficient AI in healthcare for rare diseases.
• This framework advances neuroimaging analysis and supports resource-constrained diagnostic settings.