Breast Cancer Classification with a Custom MLP

This repository contains a Python implementation of a Multi-layer Perceptron (MLP) for binary classification of breast tumors (malignant vs. benign). The model is built from scratch with configurable parameters and training options, and it also includes a comparison with scikit-learn's MLPClassifier.

Overview

This project provides functionality to:

• Data Splitting and Preprocessing:

  • Standardizes the dataset.
  • Converts tumor labels (M for malignant and B for benign) into binary values (1 and 0).
  • Splits the data into 75% training and 25% validation sets.

• Training a Custom Neural Network:

  • Configurable multi-layer perceptron with:
    • Hidden Layers: Default is 2 (configurable).
    • Hidden Nodes: Default is 20 per layer.
    • Output Nodes: 2 for binary classification.
    • Activation Functions: ReLU or Sigmoid.
    • Weight Initialization: Xavier or HeUniform.
    • Loss Function: Categorical Cross Entropy.
    • Optimizer: Adam with L2 regularization.

• Prediction and Evaluation:

  • The trained model classifies tumors and evaluates performance using key metrics:
    • Loss
    • Accuracy
    • Precision
    • Recall
    • F1-score

• Comparison with scikit-learn:

  • Uses MLPClassifier from scikit-learn for performance benchmarking.

Dataset

The dataset consists of 567 tumor samples, each represented by multiple feature measurements. The labels are:

• M: Malignant
• B: Benign

Requirements

• Python 3.6+
• Required dependencies (install via pip):
  • numpy
  • pandas
  • scikit-learn
  • matplotlib
  • tqdm

To install all requirements, run:

pip install -r requirements.txt

Usage

The program supports several modes of operation via command-line arguments:

split

Splits and preprocesses the dataset into training and validation sets.

python3 ./src/main.py split

train

Trains the custom neural network with the following default parameters:

• Hidden Layers: 2
• Hidden Nodes: 20 (configurable)
• Weight Initialization: He (default) or Xavier
• Activation Function: ReLU (default) or Sigmoid
• Optimizer: Adam with L2 regularization

python3 ./src/main.py train

Additional options, such as disabling the progress bar or logging, are available via command-line flags.

predict

Uses a trained model (saved weights and biases) to classify validation data and evaluate performance.

python3 ./src/main.py predict

sklearn

Trains and evaluates an MLPClassifier from scikit-learn for performance comparison.

python3 ./src/main.py sklearn

Configuration

You can modify the following parameters directly in the code or extend the argument parser for runtime adjustments:

Network Architecture

• Hidden Layers: Default is 2.
• Hidden Nodes: Default is 20.
• Output Nodes: Default is 2.

Activation Functions

• Options: relu or sigmoid

Weight Initialization

• Options: xavier or he

Training Parameters

• Epochs, batch size, learning rate, early stopping patience, and L2 regularization strength.

Logging and Plotting

Logging

• Training and evaluation details are logged in a file (.log).
• Use --disable-logs to disable logging.

Plotting

• Training and validation loss/accuracy curves are plotted.
• Use --disable-plot to disable plotting.

License

This project is licensed under the MIT License.

Acknowledgements

• The dataset used in this project is assumed to be derived from a publicly available breast cancer dataset.
• This project compares the custom MLP model with scikit-learn's MLPClassifier to evaluate performance and flexibility.