AboutBackgroundAnalysisModelImagesResults
Background

Reading nuclei
as evidence.

Breast cancer is the most commonly diagnosed cancer worldwide; early cytological screening through fine-needle aspirate (FNA) imaging has been an important first line of triage since the early 1990s. The Wisconsin Diagnostic Breast Cancer (WDBC) dataset, collected at UW-Madison, encodes 30 morphometric features computed from digitised FNA images into a structured tabular form.

Our project re-examines this dataset through the lens of support vector machines. We compare four kernels — linear, polynomial, RBF and sigmoid — under a unified feature-selection and validation pipeline, then expose the model behaviour interactively so that clinicians and students alike can build intuition.

The model is not a diagnostic tool. It is a structured perspective on a classic dataset and a vehicle for explainable-AI experimentation.

Motivation in numbers

569
Patient samples
Wisconsin Diagnostic
30
Source features
10 mean · 10 SE · 10 worst
62.7%
Benign prevalence
357 / 212 split
1990
Original publication
Wolberg & Mangasarian
Exploratory data analysis

The signal hiding in the histograms.

Each cell below overlays the benign and malignant distributions for one of the ten mean-summary features. Hover a card to inspect it; the focused view on the right exposes the full distribution and the per-class means.

concave_points_mean

Concave points (mean)

Number of concave portions of the contour.

Benign μ ± σ
0.026 ± 0.016
Malignant μ ± σ
0.088 ± 0.035
benignmalignantdashed = per-class mean
Feature selection

Recursive feature elimination.

We ran 10-fold RFE with a linear SVM as the base estimator and tracked the drop in cross-validated F1 as features were eliminated. Six features survive the 0.05 significance cut.

01
Radius (mean)
radius_mean
importance 92%
rank 1
keep
02
Perimeter (mean)
perimeter_mean
importance 90%
rank 2
keep
03
Area (mean)
area_mean
importance 88%
rank 3
keep
04
Concavity (mean)
concavity_mean
importance 81%
rank 4
keep
05
Concave points (mean)
concave_points_mean
importance 78%
rank 5
keep
06
Texture (mean)
texture_mean
importance 41%
rank 6
keep
07
Compactness (mean)
compactness_mean
importance 38%
rank 7
drop
08
Symmetry (mean)
symmetry_mean
importance 27%
rank 8
drop
09
Smoothness (mean)
smoothness_mean
importance 23%
rank 9
drop
10
Fractal dim. (mean)
fractal_dimension_mean
importance 9%
rank 10
drop
Analysis

Choose a lens.
Same data, different posture.

Global SHAP attributions

mean(|SHAP|) · RBF kernel

Negative bars pull the prediction toward benign, positive bars toward malignant. Concave-points, perimeter, and radius dominate the global ranking — consistent with the original Wolberg/Mangasarian findings.

Misclassification risk zones

decision vs. confidence
benign — correctmalignant — correctmisclassified

LIME — local explanations

Patient #042
Predicted malignant
confidence
97%
concave_points_mean
+0.34
perimeter_mean
+0.27
radius_mean
+0.21
area_mean
+0.15
smoothness_mean
-0.05
Patient #137
Predicted benign
confidence
92%
radius_mean
-0.28
concave_points_mean
-0.22
compactness_mean
-0.12
texture_mean
+0.07
symmetry_mean
-0.05
  1. [01]
    Multisurface method of pattern separation for medical diagnosis applied to breast cytology
    W. Wolberg, O. Mangasarian · Proceedings of the National Academy of Sciences · 1990
  2. [02]
    Wisconsin Diagnostic Breast Cancer (WDBC) Dataset
    W. Street, W. Wolberg, O. Mangasarian · UCI Machine Learning Repository · 1995
  3. [03]
    Support-vector networks
    C. Cortes, V. Vapnik · Machine Learning, 20(3) · 1995
  4. [04]
    Gene selection for cancer classification using support vector machines
    I. Guyon, J. Weston, S. Barnhill, V. Vapnik · Machine Learning, 46 · 2002
  5. [05]
    A unified approach to interpreting model predictions
    S. Lundberg, S.-I. Lee · Advances in Neural Information Processing Systems · 2017
  6. [06]
    “Why should I trust you?”: Explaining the predictions of any classifier
    M. Ribeiro, S. Singh, C. Guestrin · KDD · 2016
  7. [07]
    TumorSense — DS3 Spring Project (internal report)
    N. Trueba, K. Shah, S. Ngo, E. Park · Data Science Student Society @ UC San Diego · 2026