Tumor Sense is an end-to-end SVM classification pipeline built on the Wisconsin Breast Cancer Diagnostic dataset. It reduces 30 raw cell nucleus measurements down to the 10 most predictive features and achieves high discriminatory performance providing a reliable, data-driven second opinion for ambiguous fine needle aspiration (FNA) cases.
Tumor size alone leaves a large overlap zone where benign and malignant cases are indistinguishable. Tumor Sense uses 10 cell nucleus measurements together to resolve exactly these ambiguous cases.
| Metric | Score |
|---|---|
| Malignant Recall | 98% |
| Malignant Precision | 98% |
| ROC-AUC | 0.995 |
A false negative (calling a malignant tumor benign) is a life-threatening error. The model was evaluated with this asymmetry in mind.
- Dataset: UCI Wisconsin Breast Cancer Diagnostic Dataset (569 samples, 30 features)
- Stratified train-test split to preserve class balance across both sets
- Applied Recursive Feature Elimination (RFE) with a
LinearSVCestimator - Reduced 30 raw features → top 10 most predictive features
- Algorithm: Support Vector Machine with RBF kernel
- Hyperparameter tuning:
GridSearchCVoverCandgamma - Best parameters selected via cross-validated grid search
- Precision, Recall, F1 per class
- ROC-AUC score
- SHAP summary plot for model explainability
texture3 and concave_points1 carry the most weight in pushing predictions toward malignant. High feature values (pink) generally shift the model toward malignant; low values (blue) toward benign.
tumor-sense/
├── data/
│ └── wdbc.csv # Wisconsin Diagnostic Breast Cancer dataset
├── notebook/
│ └── tumor_sense.ipynb # Full analysis and model pipeline
├── models/
│ └── svm_model.pkl # Serialized trained model
├── visuals/
│ ├── misclass.png # misclassification zone visualization
│ ├── rfe_feature_rankings.png # RFE features ranked
│ └── shap.png # SHAP feature importance plot
└── README.md
# Clone the repo
git clone https://github.com/nathaniel-trueba/tumor-sense.git
cd tumor-sense
# Install dependencies
pip install -r requirements.txt
# Launch the notebook
jupyter notebook notebook/tumor_sense.ipynb- Vedant Vardhaan - Mentor
- Nathaniel Trueba - Group Member
- Kavya Shah - Group Member
- Evan Park - Group Member
- Steven Ngo - Group Member
Tumor Sense is an academic project and is not intended for clinical use. All results are derived from a publicly available research dataset and should not be used to inform medical decisions.