DRI-NET — brain tumor classification from MRI with a deep stack ensemble

About the project

Brain tumors are hard to classify from MRI because scans vary between machines, and many models that do well on one dataset fail on another. This research merges three public MRI datasets into one larger, more varied dataset and trains a deep stack ensemble called DRI-NET on it.

The ensemble combines the strengths of VGG16, ResNet50 and EfficientNet-B7 and reached 94.33% accuracy, an F1-score of 94.27% and an MCC of 92.49%, beating every single model on its own. The work was written up as a 6-page paper in the IEEE conference format.

Key features

• Deep stack ensemble

  VGG16, ResNet50 and EfficientNet-B7 combined into one stronger model.

• 4 classes

  Glioma, meningioma, pituitary tumor or healthy brain.

• Multi-source dataset

  Three public MRI datasets merged for more varied training data.

• Strong results

  94.33% accuracy, 94.27% F1-score and 92.49% MCC.

• Better generalisation

  Works across scanners and image styles, not just one dataset.

• IEEE-format paper

  A 6-page conference paper written in LaTeX, ready to submit.

How it works

1. 1Merge 3 MRI datasets
3. 2Preprocess images
5. 33 CNNs learn features
7. 4Ensemble decides
9. 5Tumor type shown

Tech stack

Deep learning

VGG16ResNet50EfficientNet-B7Transfer learningStacked ensemble

Data

3 public MRI datasetsImage preprocessing

Write-up

LaTeXIEEE templateOverleaf

What you get with a project like this

• Full source code & notebooks
• Trained ensemble model
• Accuracy, F1 & MCC results with charts
• Paper in IEEE or your journal format
• Slides to present your paper
• One-to-one walkthrough of every step