This project presents an advanced medical image analysis system for automated brain tumor classification from MRI scans. Developed as a Final Group Project for the Machine Learning and Data Visualization course, the system integrates Transfer Learning (CNNs) with Bio-inspired Metaheuristic Optimization to achieve superior feature selection and diagnostic accuracy.
The full end-to-end pipeline—from preprocessing to GWO feature selection—is available as an interactive notebook: 👉 Open in Google Colab
The system leverages a hybrid approach to medical classification:
- Custom CNN: Baseline model built from scratch.
- Transfer Learning Backbones: VGG16, ResNet50, and DenseNet121 used for robust feature encoding.
Instead of using raw high-dimensional features, this project employs two specialized optimization techniques to reduce noise and improve SVM classification:
- Grey Wolf Optimizer (GWO): Applied to DenseNet121 features to find the optimal diagnostic subset.
- Morphogenetic Algorithm: An experimental bio-inspired method mimicking cellular differentiation for adaptive feature evolution.
- Domain: Healthcare — Medical Image Analysis (Brain MRI).
- Classes: Glioma, Meningioma, Pituitary Tumor, and No Tumor.
-
Dynamic Preprocessing: Images resized to 224 × 224 with on-the-fly normalization (
$1/255$ ) via KerasImageDataGenerator.
The project includes deep analytical insights communicated through static visualizations:
- Convergence Analysis: Tracking GWO stability and feature reduction.
- Performance Benchmarks: Accuracy vs. Computational Cost scatter plots.
- Error Analysis: Confusion Matrix heatmaps for multi-class diagnostic precision.
├── Final_Report.pdf # Full academic analysis and results
├── Code.pdf # Exported execution log from Google Colab
├── Presentation.pptx # High-level executive summary
├── Images/ # Accuracy plots, loss curves, and GWO results
└── README.md
Developed by Khaled Walid