Breast Cancer Detection using Machine Learning (Wisconsin Diagnostic)

Breast cancer is among the most common cancers worldwide, and early classification of suspicious tissue as benign or malignant changes treatment paths completely. The Wisconsin Diagnostic Breast Cancer (WDBC) dataset — 569 digitised fine-needle-aspirate samples, each described by 30 computed nuclear features — is the standard teaching benchmark for this task: clean, well-documented, and small enough to train on a laptop. Students often treat it as a generic classification exercise and miss what makes it interesting: the class imbalance is mild but the cost of errors is not symmetric — a false negative (missing a malignancy) is far worse than a false positive, which is why sensitivity and the confusion matrix matter more here than headline accuracy. This project trains a proper pipeline with stratified splitting and cross-validation, evaluates with the clinically relevant metrics, and ships a risk-predictor web demo. Every figure on the page is either from the dataset documentation or labelled as a design target — nothing is presented as a measured result before the training run.

1. The 569 WDBC samples are loaded and split 80/20 with stratification, preserving the benign/malignant ratio in both sets.
2. Features are standardised (zero mean, unit variance) with the scaler fit on the training split only.
3. A Random Forest (200 trees) is trained; 5-fold cross-validation checks hyperparameter stability before final evaluation.
4. The held-out test set (114 samples) is evaluated once: accuracy, sensitivity, specificity, ROC-AUC and the confusion matrix are recorded for the report.
5. Feature importances are extracted and plotted, giving the viva a concrete "what does the model look at" answer.
6. The web demo embeds a simplified scoring function over the same feature set for interactive demonstration; the notebook's trained model is the deliverable, not the demo's approximation.

Tech stack:

• Python 3, scikit-learn (Random Forest, metrics, CV)
• NumPy, pandas, Matplotlib
• Jupyter Notebook (training & evaluation)
• HTML5 + JavaScript (risk-predictor demo)
• WDBC dataset (UCI ML Repository)

Dataset & model details

• Dataset: Wisconsin Diagnostic Breast Cancer (WDBC) — 569 FNA samples, 30 numeric features (mean, s.e., worst of 10 nuclear properties: radius, texture, perimeter, area, smoothness, compactness, concavity, concave points, symmetry, fractal dimension), 2 classes: 357 benign (62.7%), 212 malignant (37.3%). Source: UCI ML Repository; Wolberg, Street & Mangasarian, University of Wisconsin.
• Task: Binary classification; input = 30 standardized nuclear features, output = benign/malignant probability.
• Model: Random Forest, 200 trees, trained on standardized features after a stratified 80/20 split with 5-fold cross-validation for hyperparameter checking.
• Metrics: Accuracy 97.4% (design target), sensitivity 97.7% (design target), specificity 97.2% (design target), ROC-AUC 0.99 (design target), confusion matrix on the 114-sample held-out test set. No metric is claimed as measured until the training run is executed for the order.

• [Real WDBC dataset] 569 samples, 30 features (mean, standard error and worst of 10 nuclear properties), 357 benign / 212 malignant — sourced from the UCI ML Repository.
• [Full training notebook] Loading, stratified 80/20 split, standardisation, Random Forest training, 5-fold cross-validation and held-out evaluation in one reproducible notebook.
• [Clinically relevant metrics] Accuracy plus sensitivity, specificity, ROC-AUC and the confusion matrix — with the false-negative cost discussed explicitly.
• [Risk-predictor web demo] Enter the nuclear measurements via sliders and get a live risk score with a gauge and a plain-language verdict.
• [Feature importance analysis] Random Forest importances identifying the most predictive nuclear properties (typically worst concave points, worst perimeter).
• [Class-imbalance handling] Stratified splits preserve the 62.7/37.3 ratio; the report discusses why accuracy alone would mislead.
• [Clinical-limits documentation] Explicit educational-prototype framing: not a medical device, dataset biases, and why deployment needs clinical validation.

• Complete training & evaluation Jupyter notebook
• Trained Random Forest pipeline (.pkl) with preprocessing
• Risk-predictor web demo (sliders → live risk gauge)
• Confusion matrix, ROC curve and feature-importance plots
• Project report PDF (dataset background, methodology, metrics, clinical limits)
• PPT presentation for final review
• Viva Q&A preparation document (Random Forest, sensitivity vs accuracy, overfitting, clinical validation)

• Educational prototype only — not a medical device; must never be used for diagnosis or screening decisions.
• Trained on 569 digitised FNA samples from one institution — it does not generalize to other imaging or populations without revalidation.
• The demo's live scorer is a simplified approximation for illustration; the notebook's trained model is the real deliverable.
• 97.4% accuracy is a design target, stated honestly — the report documents the actual achieved figures after training.
• No patient data is collected or stored by the demo; all inputs stay in the browser.