BreastCancer

Breast Cancer Tabular Data Generation

This project is the first of two sections related to the generation of synthetic medical data for applications in Machine Learning and Data Science. The generation of medical data presents peculiar and delicate critical issues that must be handled with caution. The goal of this project is to train a Wasserstein Generative Adversarial Network (WGAN) capable of capturing and reproducing features of the UCI Machine Learning Breast Cancer dataset such that the data generated:

• are as similar as possible to the real thing
• are indistinguishable for a classifier trained on the real data (its performance is equal on both datasets)

Classifier comparison

---

Table of Contents

1. Libraries
2. Data Preparation
3. Data Modeling
4. Explicit Evaluation
5. Latent Evaluation
6. Installation
7. Usage
8. Results
9. Conclusion
10. References

---

Libraries

The project uses the following Python libraries:

• PyTorch for model implementation.
• Pandas/NumPy for data manipulation.
• Scikit-learn for metrics and preprocessing.
• Matplotlib/Seaborn for visualization.
• SciPy for statistical tests (Kolmogorov-Smirnov, Wasserstein distance).

import pandas as pd
import numpy as np
from scipy import stats
import matplotlib.pyplot as plt
import seaborn as sns
import torch

---

Data Preparation

The dataset contains 10 features related to cancer cell measurements:

1. Radius: value of the average distance from the center to the points on the core perimeter.
2. Texture: variation of intensity in adjacent pixels.
3. Perimeter: total length of the core contour.
4. Area: area occupied by the cell nucleus.
5. Complexity: uniformity of the contours of the nucleus.
6. Compacteness: relationship between area and perimeter of the core.
7. Concavity: gravity of indentations along the perimeter of the core.
8. Number of concavities: number of concave points along the core boundary.
9. Symmetry: similarity between the two halves of the core.
10. Fractal Dimension: measure of the complexity of the shape of the core based on the fractal dimension.

Mean, standard deviation, and worst measure are reported for each feature. Of the 569 measurements 357 are of benign tumors and 212 of malignant tumors.

The entire dataset was loaded and preprocessed:

• numerical features were scaled between [-1, 1] to promote model convergence via the hyperbolic tangent function.
• class labels were encoded and provided to the model to condition the generated data.

---

Data Modeling

The two models of the WGAN were created following the best practices of the case and seeking some consistency with the number of instances in the dataset (17070):

• Generator: 23.646k parameters
• Critic: 26.753k parameters

In this way we were able to:

• adequately capture the complex and interconnected features of the dataset
• avoid overfitting by having slightly more parameters than the number of cases
• obtain quality gradients due to a more powerful generator critic

The Kolomonorv-Smirnov test was used to evaluate the null hypothesis that the distributions of each real and synthetic feature were from the same population and the Wasserstein distance to calculate the mathematical distance between each real-fact pair of distributions:

• KS test: for obi feature recreated the null hypothesis was not refutable, all features generated come from the same population as the real ones.
• W distance: for 5 features the distance from the corresponding real distributions was found to be greater than 1.

Feature	W Distance
mean perimeter	1.236084
worst perimeter	1.615902
area error	2.506679
mean area	16.915316
worst area	29.057681

These features relate to characteristics that can be technically derived through the geometric formulas of the circumference and area of the circle. Through empirical tests, an attempt was made to have the model generate the Δ between the actual and theoretical values obtained by approximating each cell to a circle: the results were not promising so a more classical approach was preferred. Nevertheless, through a graphical analysis the critical features seem to adequately trace the distributions of the actuals.

Critic features boxplots

Although only 84% of the dataset was adequately recreated the global and class-specific linear relationships were adequately reproduced:

Correlation	Mean Absolute Difference
Full Overall	0.0423
Full Benign	0.0587
Full Malignant	0.0482
Critic Overall	0.0156
Critic Benign	0.0183
Critic Malignant	0.0646

---

Explicit Evaluation

Through a simulation of an operational context, i.e., the generation of 1000 datasets, the model was evaluated in its ability to adequately reflect reality in the absence of seeds for reproducibility:

• for all datasets the null hypothesis of the KS test was not rejected for any feature
• the number of features with Wasserstein distance greater than 1 remained consistent across generations with a slight downward trend

Number of feature	Number of Datasets
4	25
5	974
6	1

To asses the quality of the data, the Train on Real Test on Synthetic (TOR-TS) and Train on Synthetic Test on Real (TOS-TR) protocol was used:

• a real and synthetic dataset were divided into train (80%), validation (10%) and test (10%) sets.
• two binary, 12.354k parameters classifiers via 128-case batches were trained on the two respective datasets showing acceptable performance, defined by both accuracy, precision, recall and F1 and a False Negative rate as low as possible on the test set.
• finally, both models were tested on the entire dataset not used for their training showing, once again, acceptable performance

TOR-TS confusion matrix (on the left), TOS-TR confusion matrix (on the right)

While the performance is adequate, the real-world impact of the False Negative rate should never be underestimated: errors of this kind can potentially lead to underestimation. The data, even if of excellent quality, and the conclusions reached by any model must always be backed up by a human decision.

---

Latent Evaluation

Since the dataste is very complex, an analysis of not only explicit but also latent similarity was necessary, carried out through Principal Component Analysis, a dimensional reduction technique that linearly combines real variables into latent components.

Explained Variance

The variance explained by the real and synthetic components appears to be virtually the same, especially for the former components, which in both datasets explain almost all of the variance.

Real EV	Fake EV
0.982045	0.988461
0.016176	0.010266
0.001558	0.001033
0.000121	0.000139
0.000088	0.000085

Cosine Distance

The cosine distance between each pair of components was calculated to check their alignment in space: the first components are more closely aligned suggesting greater similarity.

Component	Cosine Distance
1	0.0000
2	0.0014
3	0.0030
4	1.9169
5	1.9158

Wasserstein Distance

Once again, the prince tool confirmed to us the similarity between real and synthetic data.

Component	Wasserstein Distance
1	0.049311
2	0.131439
3	0.082486
4	0.174529
5	0.135763

Screeplot

Through the screeplot we see how the real and synthetic components are almost completely overlapping. Components' Screeplot

Bivariate KDE plot

The bivariate KDE plot also shows good overlap for the first components, indicating good data quality at the latent level. Bivaraite KDE

---

Installation

To use this script, follow the steps below:

1. Clone the repository
   Open the terminal and clone the GitHub repository with the command:

   git clone https://github.com/Sim98B/TabularDataGeneration.git
   cd TabularDataGeneration/BrestCancer/Model

2. Install requirements

   pip install -r requirements.txt

---

Usage

To generate a synthetic dataset based on the Breast Cancer dataset, use the following command:

python GenerateBreastCancerData.py --weights BCWGAN_weights.pth --scaler min_max_scaler.pkl

Required Arguments:

• --weights: Path to the generator weights file (.pth).
• --scaler: Path to the saved scaler file (.pkl).

Options:

• --benign: Number of samples to generate for the benign class (default: 357).
• --malignant: Number of samples to generate for the malignant class (default: 212).
• --seed: Set a seed for reproducibility.
• --output: Name of the output CSV file (default: SyntheticBreastCancer.csv).

---

Results

• Pre-trained model to generate data is saved in the Model/ directory.
• Full notebook analysis is stored the WGANBreastCancer.ipynb.
• All functions used in the implementation of the project are in the utils.py file.

---

Conclusion

The Wasserstein cGAN successfully generated synthetic breast cancer data, preserving the statistical properties of the original dataset. The evaluation through statistical measures (such as the inverse transformation of generated data using MinMax scaling) and practical validation (by ensuring class separation between benign and malignant cases) suggests that the synthetic samples closely resemble real-world distributions. This approach demonstrates its potential for augmenting medical datasets, particularly in scenarios where data privacy and scarcity are critical concerns. With further refinement, including adversarial validation and classifier-based assessments, this method could enhance data availability for machine learning applications in oncology and other medical domains.

---

References

• PyTorch Documentation
• Breast Cancer Dataset: UCI Machine Learning Repository
• Arjovsky, M., et al. (2017). “Wasserstein GAN.” ICML.

THIS PROJECT IS SOLELY INTENDED TO DEMONSTRATE HOW DEEP LEARNING AND GENERATIVE AI MODELS CAN POTENTIALLY CONTRIBUTE TO THE MEDICAL FIELD. THE GENERATED DATA ARE NOT MEANT TO SUBSTITUTE REAL DATA BUT RATHER TO SERVE AS AN ADDITIONAL SUPPORT. ANY DECISION BASED ON THESE DATA SHOULD NOT REPLACE CAREFUL HUMAN ANALYSIS.