Breast Cancer Predictor

Neural network classifier for breast mass malignancy prediction — interactive Streamlit UI with real-time radar chart visualization.
PyTorch · Streamlit · Plotly · scikit-learn · Wisconsin Breast Cancer Dataset

---

What This Is

A fully interactive breast cancer prediction app built on a custom PyTorch neural network trained on the Wisconsin Breast Cancer Dataset (UCI). Users adjust 30 biopsy measurements via sliders and get an instant benign/malignant prediction with probability scores — visualized on a live radar chart.

Built to demonstrate end-to-end ML: data cleaning → model architecture → training → serialization → interactive deployment.

---

Model Performance

Metric	Value
Test Accuracy	97.37%
Training Accuracy	97.58%
Precision (Benign)	0.97
Recall (Benign)	0.99
F1 (Benign)	0.98
Precision (Malignant)	0.98
Recall (Malignant)	0.95
F1 (Malignant)	0.97
Weighted avg F1	0.97
Test set size	114 samples

Training converged smoothly from 87.9% accuracy at epoch 10 to 97.6% at epoch 100 with BCELoss dropping from 0.52 → 0.11.

---

How It Works

Wisconsin Breast Cancer Dataset (569 samples · 30 features)
│
▼
┌─────────────────────┐
│ Data Preprocessing │ Drop ID/unnamed cols · encode M→1, B→0
│ StandardScaler │ 80/20 train/test split · feature scaling
└────────┬────────────┘
│
▼
┌─────────────────────┐
│ Neural Network │ Input(30) → Linear → ReLU
│ (PyTorch) │ → Linear(54) → Sigmoid → Output(1)
│ │ BCELoss · Adam(lr=0.001) · 100 epochs
└────────┬────────────┘
│
▼
┌─────────────────────┐
│ Model Persistence │ pickle → model.pkl + scaler.pkl
└────────┬────────────┘
│
▼
┌─────────────────────┐
│ Streamlit App │ 30 sliders · Plotly radar chart
│ │ Real-time prediction + probability scores
└─────────────────────┘

---

Model Architecture

NeuralNet(
(fc1): Linear(in=30, out=54)
(relu): ReLU()
(fc2): Linear(in=54, out=1)
(sigmoid): Sigmoid()
)

Hyperparameter	Value
Input features	30 (cell nucleus measurements)
Hidden layer size	54
Output	1 (sigmoid — malignancy probability)
Loss function	Binary Cross Entropy (BCELoss)
Optimizer	Adam (lr = 0.001)
Epochs	100
Train/test split	80/20 · random_state=32

---

Dataset

Wisconsin Breast Cancer Dataset — 569 samples, 30 numerical features computed from digitized FNA (fine needle aspirate) images of breast masses.

Features are grouped into 3 measurement types across 10 cell nucleus characteristics:

Group	Features
Mean	radius, texture, perimeter, area, smoothness, compactness, concavity, concave points, symmetry, fractal dimension
Standard Error	Same 10 characteristics
Worst	Same 10 characteristics (largest mean of 3 worst values)

Class distribution: 357 Benign (62.7%) · 212 Malignant (37.3%)

---

App Features

• 30 interactive sliders — one per feature, min/max derived from dataset range, default at feature mean
• Live radar chart — three overlapping traces (Mean / Standard Error / Worst) update with every slider change using Plotly
• Real-time prediction — benign or malignant classification with probability score for each class
• Min-max normalization — slider values normalized to [0,1] before inference to match training distribution
• GPU/CPU auto-detection — runs on CUDA if available, falls back to CPU

---

Quick Start

# Clone the repo
git clone https://github.com/DebugJedi/CancerPrediction.git
cd CancerPrediction
# Install dependencies
pip install -r requirements.txt
# Run the app
streamlit run cancer_prediction.py
# → http://localhost:8501

---

Train the Model Yourself

python model/main.py

This will:

1. Load and clean Datasets/data.csv
2. Scale features with StandardScaler
3. Train for 100 epochs, printing loss + accuracy every 10 epochs
4. Print test accuracy and full classification report
5. Save model.pkl and scaler.pkl to resources/

---

Project Structure

CancerPrediction/
├── cancer_prediction.py ← Streamlit app · UI · radar chart · prediction
├── model/
│ └── main.py ← NeuralNet class · training · evaluation · save
├── Datasets/
│ └── data.csv ← Wisconsin Breast Cancer Dataset
├── resources/
│ ├── model.pkl ← Trained PyTorch model (serialized)
│ └── scaler.pkl ← Fitted StandardScaler
├── assets/
│ └── style.css ← Custom Streamlit styling
├── .streamlit/ ← Streamlit config
└── requirements.txt

---

Tech Stack

Component	Technology
Neural network	PyTorch (nn.Module)
Data processing	pandas · scikit-learn (StandardScaler, train_test_split)
Visualization	Plotly (Scatterpolar radar chart)
UI framework	Streamlit
Model persistence	pickle
Evaluation	scikit-learn (accuracy_score, classification_report)

---

Roadmap

• Custom PyTorch neural network
• Interactive Streamlit UI with 30 feature sliders
• Real-time radar chart visualization
• Probability scores for each class
• Model and scaler persistence
• Cross-validation and hyperparameter tuning
• ROC curve and AUC visualization
• SHAP feature importance explanations
• Streamlit Cloud deployment

---

Author

Built and maintained by Priyank Rao — Data Scientist / ML Engineer
Portfolio · GitHub

---

Disclaimer: This app is for educational and research purposes only. It is not a substitute for professional medical diagnosis.