MobileNetV3 - Brain Tumor Classification on MRI Scans

This repository contains code for training and evaluating a MobileNetV3-based model to classify brain MRI images according to the presence or absence of tumors.

Objective

Train a binary image classification model using MobileNetV3 to distinguish between:

• MRI images with brain tumors
• MRI images without brain tumors

Dataset

• Source: Kaggle - Brain Cancer MRI Dataset
• Classes: 'brain_glioma', 'brain_menin', 'brain_tumor'
• Image format: .jpg
• Input size: 224x224 pixels

Preprocessing

• Resize images to 224x224
• Normalize pixel values
• Split into training and validation sets
• Data augmentation using ImageDataGenerator

Model

• Base architecture: MobileNetV3Large pre-trained on ImageNet
• Custom head:
  • GlobalAveragePooling2D
  • Dropout (rate = 0.2)
  • Dense layer with 1 unit and sigmoid activation
• Optimizer: Adam
• Loss function: Binary Crossentropy
• Metric: Accuracy

Training

• Epochs: 10
• Batch size: 32
• EarlyStopping monitoring val_loss with patience of 3 epochs

Results

• Best validation accuracy: 95.71% at epoch 9
• Best validation loss: 0.1289 at epoch 9
• Accuracy gap (train vs validation): 11.97%

Requirements

• Python ≥ 3.7
• TensorFlow ≥ 2.x
• matplotlib
• numpy
• scikit-learn

Install dependencies:

pip install -r requirements.txt

Running

To train the model:

python MobileNetV3.ipynb

Or open the notebook using Jupyter:

jupyter notebook MobileNetV3.ipynb

Repository Structure

├── MobileNetV3.ipynb         # Main notebook
├── /dataset/                 # Image dataset (not included)
├── /plots/                   # Accuracy and loss charts (optional)
├── README.md
└── requirements.txt          # Optional dependencies file

License

This project is licensed under the MIT License.