Efficient attention-based Ghost-ResNet for brain tumor classification in magnetic resonance imaging (MRI)

Area of Science:

• Medical Image Analysis
• Artificial Intelligence
• Computer Vision

Background:

• Brain tumor classification from MRI is challenging, especially under computational constraints.
• Existing deep learning models often require significant resources, limiting clinical deployment.
• There is a need for efficient yet accurate models for practical medical imaging applications.

Purpose of the Study:

• To propose an efficiency-oriented deep learning architecture for brain tumor classification.
• To enhance feature learning and reduce computational overhead in medical image analysis.
• To balance classification accuracy with computational efficiency for clinical use.

Main Methods:

• Developed a lightweight deep learning architecture integrating Ghost modules with a ResNet-50 backbone.
• Incorporated Efficient Channel Attention (ECA) blocks to improve feature discriminability.
• Evaluated the model on the Bangladesh Brain Cancer MRI Dataset (6,056 images) with CLAHE preprocessing and selective data augmentation.

Main Results:

• Achieved an overall classification accuracy of 97.85% for glioma, meningioma, and pituitary tumors.
• Macro-averaged precision, recall, and specificity all exceeded 97.8%.
• Demonstrated a 1.65% absolute accuracy improvement over DenseNet121 with reduced complexity.

Conclusions:

• The proposed attention-assisted lightweight architecture is effective for brain tumor classification.
• Efficiency-driven designs can achieve competitive performance without increased complexity.
• This approach is suitable for resource-constrained medical imaging applications, improving diagnostic accuracy while reducing overhead.