Use of a Multi-Layer Perceptron to Predict Malignancy in Ovarian Tumors

Herman Verrelst, Yves Moreau, Joos Vandewalle, Dirk TimmermanNeurIPS 1997

Introduction

In gynecological oncology, the preoperative phase represents a critical decision point. If an ovarian tumor is benign, the patient may undergo minimally invasive surgery or simple observation. However, if it is malignant, urgent referral to a specialized gynecologic oncologist for extensive surgery is required. This paper presents a clinical application of neural networks designed to assist surgeons in this high-stakes environment, demonstrating that a Multi-Layer Perceptron (MLP) can match the diagnostic accuracy of experienced gynecologists.

The Clinical Challenge

Traditional statistical models and standard neural network training methods often struggle with "class imbalance." In this study’s population, benign cases outnumbered malignant ones nearly 3 to 1. Furthermore, standard training metrics like Mean Squared Error (MSE) do not translate well to clinical reality. A clinician's primary concern is not an "average deviation" but rather the critical trade-off between missing a cancer (a false negative) and performing an unnecessary major surgery (a false positive).

Methodology: MLP and AUC Optimization

The researchers employed a Multi-Layer Perceptron (MLP), a classic feed-forward neural network architecture, testing two specific configurations: a 10-5-1 (10 inputs, 5 hidden neurons) and an 11-3-1 (11 inputs, 3 hidden neurons).

The novelty of their approach lies in the training strategy. They compared two methods:

  1. Standard MSE Optimization: Using a quasi-Newton method (BFGS) to minimize the squared difference between predicted and actual outcomes.
  2. AUC Maximization via Simulated Annealing: Instead of minimizing error, the team used Boltzmann Simulated Annealing to directly maximize the Area Under the Receiver Operating Characteristic Curve (AUC). This global optimization technique helps the model find the best parameters for clinical discrimination, effectively bypassing the biases introduced by class imbalance and unequal error costs.

Dataset and Features

The study utilized a dataset of 191 consecutive patients referred to the University Hospitals Leuven, Belgium, between 1994 and 1996. The cohort consisted of 140 benign tumors and 51 malignant tumors.

The model evaluated 10 to 11 indicators, including:

  • Demographics: Age and menopausal status.
  • Serum Markers: CA 125 levels. Notably, these were log-transformed to handle extreme outliers, making the data more digestible for the neural network.
  • Color Doppler Imaging (CDI): Presence of blood flow.
  • Morphologic Criteria: Ultrasound findings such as abdominal fluid, wall irregularities, and the presence of solid growths (papillations).

Results and Clinical Efficacy

The results demonstrated high discriminatory power, proving that a relatively simple neural network can perform at the level of a human expert.

  • Overall Performance: The 11-3-1 MLP, which included the CA 125 marker, achieved an impressive AUC of 97.8% for the total population.
  • Clinical Benchmarking: At a sensitivity of 98% (successfully identifying 98 out of 100 cancers), the model's false-positive rate was between 10% and 15%.
  • Human Comparison: This performance was nearly identical to the subjective impressions of two experienced gynecologists, who demonstrated a sensitivity of 98% with false-positive rates of 12% and 13%.
  • Robustness: Even without the expensive and time-consuming CA 125 test, the 10-5-1 architecture still achieved a highly respectable AUC of 96.5%.

By optimizing for AUC rather than MSE, the model successfully avoided the "prevalence trap"—where a model might achieve high accuracy simply by guessing the majority class. Furthermore, relying on ROC curves allows clinicians to select a diagnostic threshold tailored to a specific patient's risk profile.

Limitations and Future Work

Despite the strong results, the study has limitations. With only 191 patients, the dataset is small by modern machine learning standards. To prevent overtraining (memorization), the authors had to restrict the network to very few hidden neurons and employ "early stopping." Additionally, while the model processes inputs objectively, some features—like identifying "irregular walls" on an ultrasound—still rely on the subjective interpretation of a sonographer.

Future work should focus on identifying more objective indicators to reduce reliance on human interpretation during data collection. Finally, while simulated annealing proved highly effective for AUC optimization, the authors note that it remains computationally intensive.

Summary updated

BibTeX
@inproceedings{verrelst1997use,
  author = {Herman Verrelst and Yves Moreau and Joos Vandewalle and Dirk Timmerman},
  title = {{Use of a Multi-Layer Perceptron to Predict Malignancy in Ovarian Tumors}},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {1997},
}
Figure