Brain Tumor Classification using CNN

This project presents a Convolutional Neural Network (CNN)-based approach for brain tumor classification from MRI scans, implemented using Tensorflow/Keras. The model classifies grayscale MRI images into four clinically relevant categories: Glioma, Meningioma, Pituitary (tumors), and Nontumor. The primary objective is to evaluate how effectively a deep-learning model can distinguish between visually similar brain tumor types using structural MRI data.

Repository Structure

├── Brain_tumor_SamayAsubadin.py # Model training script
├── model_analysis.ipynb # Evaluation and visualization
├── images/ # Result figures
├── requirements.txt # Dependencies
├── .gitignore # Ignored files
└── README.md

The repository is structured to keep training logic in a clean, reproducible .py script, and an exploratory analysis and visualization in a separate Jupyter notebook.

Model Architecture

• Framework: TensorFlow / Keras
• Input: 224×224 grayscale MRI images
• Architecture:
  -3 convolutional blocks (Conv2D + MaxPooling).
  • Fully connected layers with dropout generalization.
• Optimizer: Adam.
• Output: Softmax layer with 4 classes.
• Loss Function: Categorical cross-entropy.
• Evaluation: Accuracy, Precision, Recall, F1-score, Confusion Matrix.

Dataset

The dataset is NOT included in this repository, but similar datasets are publicly available in Kaggle, containing labeled images for four classes: glioma, meningioma, pituitary tumor, and non-tumor. The dataset came preorganized into train and test directories, with the following class-wise distribution:

• Training Set:
  Glioma: 1,321 images
  Meningioma: 1,339 images
  Non-tumor: 1,596 images
  Pituitary: 1,457 images

-Testing Set:
Glioma: 299 images
Meningioma: 305 images
Non-tumor: 404 images
Pituitary: 299 images

In total, the dataset used contains 7,020 training images and 1,307 testing images. During training, 20% of the training set is further used as a validation subset via Keras’ ImageDataGenerator to monitor generalization performance and apply early stopping.

Expected directory structure:
data/
├── Train/
├── glioma/
├── meningioma/
├── nontumor/
└── pituitary/
└── Test/
├── glioma/
├── meningioma/
├── nontumor/
└── pituitary/

Results

Training Set Class Distribution

The training dataset exhibits a moderately imbalanced class distribution. Non-tumor images constitute the largest class, while glioma and meningioma samples are slightly underrepresented. Pituitary tumor images lie between these extremes. The imbalance is moderate enough to be handled without explicit class reweighting, allowing the model’s performance to be evaluated under practical conditions.

Additionally, random samples are visualized to qualitatively inspect image quality, contrast, and inter-class variability. This can be found in the separate model_analysis.ipynb notebook.

Classification Performance

The trained CNN achieves an overall test accuracy of 88%.
Class-wise performance is summarized below:

Confusion Matrix

The confusion matrix shows strong overall classification performance, with most predictions concentrated along the diagonal. Non-tumor and pituitary cases are classified with particularly high accuracy, while most misclassifications occur between glioma and meningioma tumors.

Predicted vs Actual Labels

Qualitative assessment of the model’s performance can be visualized in the model_analysis.ipynb notebook, and was done by comparing predicted labels against ground-truth annotations for randomly selected MRI scans. Correct predictions are highlighted in green, while misclassifications are shown in red, allowing for immediate visual identification of success and failure cases. This visualization complements quantitative metrics by exposing individual prediction behavior, supporting transparent evaluation of model reliability in clinical-like scenarios.

Training Curves

The training curves show a steady increase in training accuracy accompanied by a continuous decrease in training loss, indicating that the model is effectively learning patterns from the training data. In contrast, validation accuracy improves initially but then plateaus, while validation loss begins to increase after a few epochs. This divergence suggests the onset of overfitting, where the model continues to fit the training data more closely without achieving corresponding improvements on unseen data.

How to Run

1. Install dependencies

pip install -r requirements.txt

2. Train and evaluate the model

python Brain_tumor_SamayAsubadin.py

3. Open model_analysis.ipynb to explore dataset distribution, check random samples, visualize predicted vs. actual sample classification, confusion matrix visualization, and training curves.

Reproducibility

Due to randomness in weight initialization and data shuffling, exact results may vary slightly across runs.