A robust approach for multi-type classification of brain tumor using deep feature fusion

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Biology

Background:

• Accurate brain tumor classification is critical for effective patient treatment planning.
• Existing classification methods require enhancement in accuracy and robustness.

Purpose of the Study:

• To develop a deep feature fusion method using convolutional neural networks (CNNs) for improved brain tumor classification.
• To enhance classification accuracy and robustness while mitigating over-fitting risks.

Main Methods:

• Feature extraction using three pre-trained CNN models: ResNet101, DenseNet121, and EfficientNetB0.
• Fine-tuning models and adjusting extracted features for consistent shape.
• Pairwise summation for feature fusion, followed by classification.

Main Results:

• The proposed deep feature fusion method demonstrated superior performance on public datasets (Figshare and Kaggle).
• Fusion of ResNet101 and DenseNet121 features yielded the highest accuracy: 99.18% (Figshare) and 97.24% (Kaggle).

Conclusions:

• Deep feature fusion is an effective strategy for enhancing brain tumor classification accuracy.
• The proposed method offers a robust approach to assist clinicians in diagnosing brain tumors.
• The ResNet101 and DenseNet121 feature fusion combination shows exceptional promise for clinical application.