[Paper Review] Toward Efficient Breast Cancer Diagnosis and Survival Prediction Using L-Perceptron

Hadi Mansourifar, Weidong Shi|arXiv (Cornell University)|Nov 5, 2018

AI in cancer detection1 references4 citations

TL;DR

This paper introduces L-Perceptron, a novel perceptron variant that combines least squares classification with perceptron learning to enable flexible, non-linear decision boundaries. It achieves state-of-the-art performance—97.42% accuracy and 98.73% sensitivity—on the Wisconsin Breast Cancer dataset and 75.18% accuracy and 83.86% F1 score on Haberman’s Survival dataset, outperforming prior methods without preprocessing or feature selection.

ABSTRACT

Breast cancer is the most frequently reported cancer type among the women around the globe and beyond that it has the second highest female fatality rate among all cancer types. Despite all the progresses made in prevention and early intervention, early prognosis and survival prediction rates are still unsatisfactory. In this paper, we propose a novel type of perceptron called L-Perceptron which outperforms all the previous supervised learning methods by reaching 97.42 \% and 98.73 \% in terms of accuracy and sensitivity, respectively in Wisconsin Breast Cancer dataset. Experimental results on Haberman's Breast Cancer Survival dataset, show the superiority of proposed method by reaching 75.18 \% and 83.86 \% in terms of accuracy and F1 score, respectively. The results are the best reported ones obtained in 10-fold cross validation in absence of any preprocessing or feature selection.

Motivation & Objective

• To address the persistent challenge of low accuracy and sensitivity in early breast cancer diagnosis and survival prediction.
• To develop a new supervised learning model that improves performance on imbalanced and complex biomedical datasets.
• To eliminate the need for preprocessing or feature selection while maintaining high predictive performance.
• To validate the model on two standard breast cancer datasets: Wisconsin Breast Cancer and Haberman’s Survival.

Proposed method

• L-Perceptron uses a hybrid approach combining least squares classification and traditional perceptron learning to model non-linear relationships per feature.
• For each feature, a mathematical function (e.g., polynomial) is fitted to the data, with the function's parameters learned during training.
• The method dynamically determines the optimal polynomial order per feature within user-defined degree bounds (dlb, dub).
• The model employs a thresholded activation function to classify instances, with p1 and p2 as class-specific target values for positive and negative classes.
• Training is performed iteratively using a custom Python implementation (lp.py) with configurable hyperparameters including learning iterations and convergence thresholds.
• 10-fold cross-validation is used to ensure unbiased performance evaluation across both datasets.

Experimental results

Research questions

• RQ1 Can a perceptron-based model achieve superior accuracy and sensitivity in breast cancer diagnosis without data preprocessing?
• RQ2 How does the L-Perceptron's adaptive function fitting per feature improve generalization and reduce overfitting?
• RQ3 Does the L-Perceptron outperform established classifiers like SVM, Naive Bayes, and RBF networks on standard breast cancer datasets?
• RQ4 How does the model perform on imbalanced survival prediction tasks, such as Haberman’s dataset?

Key findings

• On the Wisconsin Breast Cancer Dataset, L-Perceptron achieved 97.42% accuracy and 98.73% sensitivity, the highest reported results without preprocessing or feature selection.
• In comparison to other models, L-Perceptron ranked first in accuracy and sensitivity, outperforming Naive Bayes (97.36% accuracy) and RBF networks (96.77% accuracy).
• On Haberman’s Survival Dataset, L-Perceptron achieved 75.18% accuracy and 83.86% F1 score, the best reported performance among all methods tested.
• The model demonstrated strong robustness in imbalanced settings, with high sensitivity (90.04%) despite low specificity (37.08%), indicating effective detection of positive survival cases.
• The L-Perceptron achieved the best F1 score on Haberman’s dataset, surpassing even MLP (72.64%) and Random Forest (80.38%).
• The model’s performance was stable across multiple metrics, with consistent top rankings in accuracy, sensitivity, and F1 score across both datasets.