CancerScan AI

A breast cancer prediction system built with Random Forest and SHAP explainability, deployed as a full-stack web application.

Live Demo · Kaggle Notebook · Dataset

---

What it does

Takes 5 tumor measurements as input and predicts whether a tumor is benign or malignant — then explains why using SHAP feature attribution, so the prediction is never just a black box.

---

Results

Metric	Score
Accuracy	97.2%
Precision	96%
Recall	97%
F1	97%
False negatives	3 / 114 test cases

Evaluated using 5-fold stratified cross-validation, not a single train/test split.

---

Stack

Layer	Technology
Model	scikit-learn Random Forest
Explainability	SHAP (TreeExplainer)
Backend	FastAPI + Uvicorn
Frontend	Vanilla HTML/CSS/JS
Hosting	Render (backend)

---

Project structure

cancer-app/
├── main.py               # FastAPI backend — /predict and /features endpoints
├── index.html            # Frontend — sliders, results, SHAP bar chart
├── model.pkl             # Trained Random Forest model
├── X_train.pkl           # Training data for SHAP background
├── feature_names.json    # Top 5 + all 15 feature names
├── requirements.txt      # Python dependencies
└── notebook/
    └── breast-cancer-prediction.ipynb

---

ML approach

Data cleaning

• Removed the id column (row identifier, not a medical feature)
• Label encoded diagnosis: B → 0, M → 1

Class imbalance

• Used class_weight='balanced' in Random Forest to prevent the model from being biased toward the majority benign class

Feature selection

• Ranked all 30 features by Random Forest importance
• Kept top 15 to remove redundant correlated columns (mean/se/worst versions of the same measure)

Evaluation

• 5-fold stratified cross-validation for reliable, stable metrics
• Confusion matrix analysis with focus on minimizing false negatives (missed malignant cases)

Explainability

• SHAP TreeExplainer generates per-prediction feature attribution
• Every prediction shows which measurements drove the result and by how much

---

API

GET /features Returns min, max and mean values for the top 5 features — used to populate slider ranges in the frontend.

POST /predict

{
  "inputs": {
    "concave points_worst": 0.14,
    "perimeter_worst": 120.0,
    "concave points_mean": 0.07,
    "radius_worst": 16.0,
    "area_worst": 900.0
  }
}

Returns:

{
  "prediction": 1,
  "label": "Malignant",
  "confidence": 94.2,
  "shap_values": [...],
  "shap_base": 0.37
}

---

Run locally

git clone https://github.com/aliiakbarkhan/cancer-app.git
cd cancer-app
pip install -r requirements.txt
uvicorn main:app --reload

Then open http://localhost:8000/app

---

Test cases

Feature	Benign case	Malignant case
Concave points worst	0.053	0.261
Perimeter worst	78.3	197.4
Concave points mean	0.020	0.147
Radius worst	12.1	29.8
Area worst	422.0	2501.0

---

Disclaimer

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

---

Ali Akbar Khan · Kaggle · GitHub