Metaheuristic-Driven Feature Selection for Brain Tumour MRI Classification

Group project for the Machine Learning module of the MSc Data Science at the University of Europe for Applied Sciences (2025-2026). It compares baseline and pre-trained deep networks for 4-class brain tumour MRI classification, and tests two add-ons: Grey Wolf Optimisation (GWO) feature selection with an SVM, and a morphogenesis-inspired label-propagation classifier.

What it does

• Loads the Kaggle Brain Tumor MRI Dataset (glioma, meningioma, no tumour, pituitary), resized to 224 x 224.
• Trains and evaluates on the held-out test set:
  • a basic fully connected ANN on raw pixels
  • a small custom CNN (2 conv blocks)
  • VGG16, ResNet50 and DenseNet121 (ImageNet weights, frozen-base training, then fine-tuning of the last layers)
• Extracts 256-d features from the fine-tuned DenseNet121, standardises them, runs a binary GWO-style search for a feature subset (RBF-SVM accuracy as fitness), and trains a final SVM on the selected features.
• Plots training curves, a model accuracy comparison chart, normalised confusion matrices (seaborn) and a PCA projection of the DenseNet121 features.
• Separate notebook: frozen ResNet50 features, k-nearest-neighbour similarity graph, and an iterative "local excitation / local inhibition" update of class activations, starting from 15% labelled training samples (weak supervision). Test samples are classified by a k-NN vote over the learned activations.

Tech stack

Python, TensorFlow / Keras, scikit-learn (SVC, NearestNeighbors, StandardScaler, PCA), NumPy, pandas, Matplotlib, seaborn. Run on Google Colab.

Repository contents

File	Description
Metaheuristic-Driven-Feature-Selection-in-Deep-Neural-Networks-for-Automated-Brain-Tumour-Detection.ipynb	Main notebook: ANN, CNN, VGG16, ResNet50, DenseNet121, DenseNet121 + GWO + SVM, evaluation and plots
Biomimetic_Brain_Tumor_MRI_Improved.ipynb	Morphogenesis-inspired self-organising classifier on ResNet50 features
Metaheuristic-Driven Feature Selection in Deep Neural Networks for Automated Brain Tumor Detection (1).pdf	Project report (15 pages)

Results

Test-set accuracy from the saved outputs of the main notebook (1,311 test images):

Model	Test accuracy	Test loss
ResNet50 (pre-trained, fine-tuned)	92.52%	0.236
VGG16 (pre-trained, fine-tuned)	92.45%	0.332
DenseNet121 (pre-trained, fine-tuned)	92.45%	0.226
Custom CNN	81.92%	0.454
Basic ANN	62.85%	1.013
DenseNet121 + GWO + SVM	29.40%	-

The GWO + SVM result is evaluated on an 80/20 split of the training-set features (915 samples), not on the test set, and is close to chance for 4 classes. In the notebook the features are extracted from a training generator with shuffle=True while the labels come from generator.classes (directory order), so features and labels are likely misaligned; this probably explains the low score more than the feature selection itself.

The morphogenesis-inspired notebook reaches 67% accuracy (macro F1 0.66) on its own 20% hold-out split of 1,143 images.

How to run

1. Open a notebook in Google Colab:
   • Main notebook
   • Morphogenesis notebook
2. Get the dataset:
   • Main notebook: upload the dataset to Google Drive so that Training/ and Testing/ are under /content/drive/MyDrive/archive (or change base_path).
   • Morphogenesis notebook: upload your own kaggle.json API token; the notebook downloads and unzips the dataset.
3. Use a GPU runtime and run all cells.

---

Developed by Khaled Walid, with Mujtaba Shaikh and Joshuva Jeemon (MSc Data Science, University of Europe for Applied Sciences)