Breast Cancer Diagnostic Support System

A clinical-grade AI decision support tool for breast cancer classification using Fine Needle Aspiration cytology measurements

---

---

Overview

A production-grade clinical decision support application that classifies breast tumors as Benign or Malignant using a Random Forest ensemble trained on the Wisconsin Diagnostic Breast Cancer (WDBC) dataset. The platform is designed for medical researchers and clinical informaticians, featuring explainability (SHAP), real-time monitoring, error analysis, and PDF reporting — all in a clean, accessible interface.

Medical Disclaimer: This application is intended for educational and research purposes only. It is not a substitute for clinical diagnosis by a qualified physician.

---

Application Walkthrough

1 · AI Analysis — Clinical Dashboard

Configure tumor measurements in the sidebar and run the ensemble model to get a real-time classification result.

---

2 · Diagnostic Result Cards

Binary classification result displayed in a high-contrast card with confidence score, malignancy threshold indicator, and probability breakdown.

---

3 · Model Monitoring — Operational Health

Real-time data drift detection with Z-score analysis per feature, a system health gauge, and prediction trend history.

---

4 · Error Analysis — ROC Curve & Confusion Matrix

Interactive ROC curve with configurable operating point, confusion matrix, and misclassification log. AUC: 0.9990.

---

5 · Research Lab — Synthetic Stress Testing

Generate synthetic clinical profiles via Gaussian multivariate sampling and project them onto the training dataset's PCA manifold.

---

6 · Dataset Explorer — Interactive Scatter Analysis

Filter samples by diagnosis and feature range, explore relationships through marginal-histogram scatter plots, and review raw sample data.

---

Architecture

System Architecture

graph TB
    subgraph UI["🖥️  Streamlit Frontend"]
        SB["Clinical Configuration\nSidebar"]
        T1["AI Analysis Tab"]
        T2["Model Monitoring Tab"]
        T3["Model Card & Ethics Tab"]
        T4["Error Analysis Tab"]
        T5["Research Lab Tab"]
        T6["Dataset Explorer Tab"]
    end

    subgraph CORE["Application Core (app/)"]
        MAIN["main.py\nOrchestrator"]
        UI_C["components/\nui · sidebar · visualizations\nethics · explorer"]
    end

    subgraph UTILS["Utility Layer (utils/)"]
        ML["model_loader.py\nRandom Forest + SHAP"]
        DATA["data_loader.py\nWDBC Dataset"]
        MON["monitoring.py\nDrift Detection"]
        INS["clinical_insights.py\nRule Engine"]
        ROB["robustness.py\nEnsemble Variance"]
        ERR["error_analysis.py\nROC · CM · FN/FP"]
        CF["counterfactuals.py\nDecision Boundary"]
        SEN["sensitivity.py\nSingle-Feature Curve"]
        SYN["synthetic_data.py\nGaussian Sampling"]
        REP["report_generator.py\nPDF Export"]
    end

    subgraph ASSETS["Assets"]
        MDL["models/\nbreast_cancer_model_v2.pkl"]
        CSV["data/\ndata.csv (WDBC)"]
        LOG["predictions.csv\nAudit Log"]
    end

    SB --> MAIN
    MAIN --> UI_C
    MAIN --> ML
    MAIN --> DATA
    MAIN --> MON
    MAIN --> INS
    MAIN --> ROB
    MAIN --> ERR
    MAIN --> CF
    MAIN --> SEN
    MAIN --> SYN
    MAIN --> REP
    ML --> MDL
    DATA --> CSV
    MON --> LOG

    style UI fill:#eff6ff,stroke:#3b82f6,color:#1e3a8a
    style CORE fill:#f0fdf4,stroke:#16a34a,color:#14532d
    style UTILS fill:#fefce8,stroke:#f59e0b,color:#78350f
    style ASSETS fill:#fdf4ff,stroke:#a855f7,color:#581c87

Loading

Prediction Pipeline

flowchart LR
    A(["FNA Measurements\n30 features"]) --> B["Feature Alignment\n& Validation"]
    B --> C["StandardScaler\nNormalization"]
    C --> D["Random Forest\nEnsemble (v2)"]
    D --> E{"Probability\n≥ Threshold?"}
    E -- Yes --> F(["🔴 Malignant"])
    E -- No  --> G(["🟢 Benign"])
    D --> H["SHAP Explainer\nFeature Attribution"]
    D --> I["Tree-Level Variance\nRobustness Score"]
    F --> J["Clinical Insights\nRule Engine"]
    G --> J
    J --> K["PDF Report\nGenerator"]

    style A fill:#eff6ff,stroke:#3b82f6,color:#1e40af
    style F fill:#fef2f2,stroke:#dc2626,color:#7f1d1d
    style G fill:#f0fdf4,stroke:#16a34a,color:#14532d
    style H fill:#fefce8,stroke:#f59e0b,color:#78350f
    style K fill:#fdf4ff,stroke:#a855f7,color:#581c87

Loading

Feature Taxonomy

mindmap
  root((WDBC\nFeatures))
    Mean Measurements
      Radius Mean
      Perimeter Mean
      Area Mean
      Compactness Mean
      Concavity Mean
      Concave Points Mean
      Smoothness Mean
      Symmetry Mean
      Fractal Dimension Mean
      Texture Mean
    Standard Error
      Radius SE
      Perimeter SE
      Area SE
      Compactness SE
      Concavity SE
      Concave Points SE
    Worst-Case Values
      Radius Worst
      Perimeter Worst
      Area Worst
      Concavity Worst
      Concave Points Worst
      Compactness Worst

Loading

Tab Workflow

journey
    title Clinical Workflow Through the Application
    section Configure
      Open sidebar: 9: Clinician
      Set malignancy threshold: 8: Clinician
      Adjust tumor measurements: 8: Clinician
    section Analyse
      Run AI Analysis: 9: Clinician
      Review classification result: 9: Clinician
      Inspect SHAP contributions: 8: Clinician
      Explore radar & PCA charts: 7: Clinician
    section Validate
      Check Model Monitoring drift: 8: Clinician
      Review Error Analysis ROC: 8: Clinician
      Read Model Card & Ethics: 7: Clinician
    section Export
      Generate clinical PDF report: 9: Clinician
      Download report: 9: Clinician

Loading

---

Feature Matrix

Feature	Description
AI Classification	Random Forest ensemble → Benign / Malignant with probability score
SHAP Explainability	Top-10 feature contributions, local instance explanation
Adjustable Threshold	Sensitivity / Specificity trade-off via live slider
Diagnostic Robustness	Tree-level variance, 95% confidence interval, ensemble agreement %
Radar Chart	Sample profile vs. population mean (normalized)
PCA Projection	2D manifold placement of current sample in training space
Clinical Insights	Rule-based flagging of high/medium/low severity features
What-If Analysis	Counterfactual search for minimum adjustments to Benign boundary
Sensitivity Analysis	Single-feature malignancy probability curve, all others constant
PDF Report	Structured clinical risk report with all key findings
Data Drift Monitoring	Z-score per feature vs. training distribution, health gauge
Prediction History	Audit log with trend charts for prediction frequency
Error Analysis	Interactive ROC, AUC, confusion matrix, FN/FP breakdown
Research Lab	Gaussian synthetic sample generation, PCA manifold overlay
Dataset Explorer	Filterable scatter plots, marginal histograms, summary statistics
Model Card & Ethics	Structured accountability document: intended use, limitations, fairness

---

Dataset

Property	Value
Name	Wisconsin Diagnostic Breast Cancer (WDBC)
Source	UCI Machine Learning Repository
Authors	Wolberg, Street & Mangasarian (1995)
Instances	569
Features	30 numeric (FNA cytology measurements)
Classes	Malignant (212) · Benign (357)
Provenance	Digitized FNA images from fine needle aspirate of breast mass

---

Model Performance

Training Pipeline Results (from Breast_Cancer_Prediction.ipynb)

Feature Selection — Univariate ANOVA F-test (Top 15)

Feature	F-Score	Feature	F-Score
concave_points_worst	964.39	concave_points_mean	861.68
perimeter_worst	897.94	radius_worst	860.78
perimeter_mean	697.24	area_worst	661.60
radius_mean	646.98	area_mean	573.06
concavity_mean	533.79	concavity_worst	436.69
compactness_mean	313.23	compactness_worst	304.34
radius_se	268.84	perimeter_se	253.90
area_se	243.65

Model Cross-Validation (5-Fold, Stratified)

Algorithm	CV Accuracy	Std Dev
Random Forest	0.9495	±0.0330
Logistic Regression	0.9407	±0.0315
Gradient Boosting	0.9319	±0.0377
SVM	0.9495	—

Best Random Forest Hyperparameters (GridSearchCV · 540 fits)

n_estimators=200, max_depth=None, min_samples_leaf=1, min_samples_split=10
Best CV score: 0.9516

Dataset Split (80/20 stratified, random_state=42)

Set	Samples	Benign	Malignant
Training	455	285	170
Test	114	72	42

Production Model Metrics

Metric	Value
Accuracy	~98.2%
AUC-ROC	0.9990
Sensitivity	~97% (at T=0.50)
Specificity	~99% (at T=0.50)
False Negatives	2 (at T=0.50, on test set)
Algorithm	Random Forest Ensemble (v2)
Input Features	15 (univariate-selected)
Preprocessing	StandardScaler normalization
Train/Test Split	80% / 20% stratified

---

Project Structure

Breast-Cancer-Prediction/
├── .github/
│   └── workflows/
│       └── ci.yml               # GitHub Actions CI/CD pipeline (6 jobs)
├── .streamlit/
│   └── config.toml              # Design token theme config
├── app/
│   ├── main.py                  # Application orchestrator & tab layout
│   ├── components/
│   │   ├── ui.py                # CSS design system & page components
│   │   ├── sidebar.py           # Clinical Configuration panel
│   │   ├── visualizations.py    # Plotly chart library
│   │   ├── ethics.py            # Model Card & Ethics document
│   │   └── explorer.py          # Dataset Explorer component
│   └── utils/
│       ├── model_loader.py      # Model & SHAP explainer loading
│       ├── data_loader.py       # WDBC dataset loader
│       ├── monitoring.py        # Drift detection & prediction log
│       ├── clinical_insights.py # Rule-based insight engine
│       ├── robustness.py        # Ensemble variance analysis
│       ├── error_analysis.py    # ROC, confusion matrix, FN/FP
│       ├── counterfactuals.py   # Decision boundary search
│       ├── sensitivity.py       # Single-feature sensitivity curves
│       ├── synthetic_data.py    # Gaussian synthetic sample generator
│       └── report_generator.py  # PDF report builder
├── tests/
│   ├── conftest.py              # Shared fixtures (Streamlit stub, WDBC data, ML objects)
│   ├── test_clinical_insights.py  # 20 tests — rule-based insight engine
│   ├── test_counterfactuals.py    # 12 tests — what-if boundary search
│   ├── test_data_pipeline.py      # 35 tests — dataset, scaler & split contracts
│   ├── test_error_analysis.py     # 20 tests — ROC, confusion matrix, FP/FN
│   ├── test_model_card.py         # 20 tests — HTML badge/insight block helpers
│   ├── test_monitoring.py         # 18 tests — drift detection & CSV logging
│   ├── test_report_generator.py   # 19 tests — PDF generation
│   ├── test_robustness.py         # 17 tests — ensemble variance & CI
│   ├── test_sensitivity.py        # 15 tests — feature sensitivity curves
│   └── test_synthetic_data.py     # 20 tests — synthetic sample generation
├── data/
│   └── data.csv                 # WDBC dataset (569 samples, 30 features)
├── models/
│   └── breast_cancer_model_v2.pkl
├── notebooks/
│   └── Breast_Cancer_Prediction.ipynb
├── screenshots/                 # UI documentation images
├── requirements.txt             # Runtime dependencies
├── requirements-dev.txt         # Test & CI dependencies
├── setup.cfg                    # pytest configuration
└── README.md

---

Testing

The project includes 201 automated tests covering all utility business logic.

Run the Test Suite

# Install test dependencies
pip install -r requirements-dev.txt

# Run all tests
pytest tests/

# Run with coverage report
pytest tests/ --cov=app --cov-report=term-missing

# Run a specific test file
pytest tests/test_monitoring.py -v

Test Coverage (Utility Layer)

Module	Coverage
clinical_insights.py	100%
sensitivity.py	100%
synthetic_data.py	100%
report_generator.py	97%
error_analysis.py	96%
counterfactuals.py	95%
monitoring.py	93%
robustness.py	91%

---

CI/CD Pipeline

GitHub Actions runs automatically on every push and pull request to main, master, and develop.

graph LR
    A[Lint] --> C[Tests]
    A --> D[Security]
    C --> E[Model Check]
    C --> F[Smoke Test]
    E --> G[Summary]
    F --> G
    D --> G

Loading

Job	Description
Lint	flake8 style enforcement on app/ and tests/
Tests	pytest matrix across Python 3.10, 3.11, 3.12 with coverage gate (≥70%)
Security	pip-audit scans requirements.txt for known CVEs
Model Check	Verifies the .pkl bundle is loadable and exposes predict_proba
Smoke Test	Imports every utility module in a headless environment
Summary	Gate job — fails pipeline if any critical stage fails

See .github/workflows/ci.yml for the full configuration.

---

Installation & Setup

1. Clone the repository

git clone https://github.com/Khanz9664/Breast-Cancer-Prediction.git
cd Breast-Cancer-Prediction

2. Create and activate a virtual environment

python -m venv venv
source venv/bin/activate        # Windows: venv\Scripts\activate

3. Install dependencies

pip install -r requirements.txt

4. Launch the application

streamlit run app/main.py

Open your browser at http://localhost:8501

---

Design System

The application uses a strict design token system for consistency across all components:

Token	Value	Usage
Primary	#3b82f6	Buttons, active tabs, focus rings
Background	#ffffff	Main workspace, cards
Secondary BG	#f8fafc	Sidebar, muted surfaces
Text	#0f172a	All body and heading text
Border	#e2e8f0	Card and input borders
Success	#16a34a	Benign classification, positive indicators
Error	#dc2626	Malignant classification, false negatives
Warning	#f59e0b	Medium-risk insights, false positives
Info	#2563eb	Informational badges, links

Typography: Inter · H1 28/700 · H2 22/600 · Body 15/400 · Labels 13/500 · Captions 12/400

---

Contributing

Pull requests are welcome. For major changes, please open an issue first to discuss the proposed modification.

---

License

This project is licensed under the MIT License — see the LICENSE file for details.

---

Author

Shahid Ul Islam Portfolio · GitHub · LinkedIn · Email

---

_{Built with Python · Streamlit · scikit-learn · SHAP · Plotly · ReportLab}