A dual neural network ensemble approach for multiclass brain tumor classification.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Neuroscience

Background:

• Accurate brain tumor classification is crucial for effective treatment planning.
• Existing computer-aided diagnosis (CAD) systems require enhancement for multiclass classification accuracy.

Purpose of the Study:

• To develop an interactive CAD system for multiclass brain tumor classification.
• To evaluate the performance of a dual-level neural network ensemble compared to a single-level network.

Main Methods:

• Extracted 218 texture and intensity features from 856 segmented regions of interest (SROIs).
• Utilized Principal Component Analysis (PCA) for feature space dimensionality reduction.
• Employed a dual-level neural network ensemble for classification of astrocytoma, glioblastoma multiforme, medulloblastoma, meningioma, metastases, and normal regions.

Main Results:

• The dual-level neural network ensemble achieved 95.85% overall accuracy, outperforming a single-level network (91.97%).
• High class-specific accuracies were reported: astrocytoma (96.29%), glioblastoma multiforme (96.15%), medulloblastoma (90%), meningioma (93.00%), metastases (96.67%), and normal regions (97.41%).
• In a stricter testing scenario (no SROIs from the same patient in training/testing), 90.4% overall accuracy was achieved.

Conclusions:

• The dual-level neural network ensemble significantly enhances brain tumor classification accuracy.
• The proposed CAD system offers a valuable tool for radiologists, improving diagnostic decision-making.
• Further validation with larger datasets and extensive training can optimize the system's performance.