PaperlayerView exact revision on arXiv

Source-linked AI summary

On Breast Cancer Detection: An Application of Machine Learning Algorithms on the Wisconsin Diagnostic Dataset

Abien Fred Agarap

arXiv:1711.07831v4cs.LGstat.ML

TL;DR

Breast cancer is a major research topic, and medical data science and machine learning may assist practitioners’ decision making. This paper compares machine learning algorithms, including GRU-SVM, for breast cancer classification using WDBC, finding high performance across the presented algorithms.

  • Problem

    Breast cancer represents 15% of new cancer cases in the United States, while growing clinical data support research into machine learning for medical decision making.

    Source: §1 · p. 1

  • Method

    The paper applies GRU-SVM, which combines a gated recurrent unit with support vector machine prediction, alongside nearest-neighbor distance methods for breast cancer classification on WDBC.

    Source: §1 · p. 1, §2.4.1 · p. 2, §2.4.4 · p. 2

  • Results

    All presented machine learning algorithms exhibited high performance on binary breast cancer classification, while L2-SVM achieved ≈96.09% test accuracy in this study.

    Source: §3 · p. 3, §4 · p. 5

  • Takeaways & Limitations

    The results support the effectiveness of machine learning algorithms for diagnosing whether a breast tumor is benign or malignant.

    Source: §4 · p. 5

    The comparison with prior SVM results is not fully fair because the studies used different training/testing partitions: 70/30 here versus 60/40 previously.

    Source: §3 · p. 3

Abstract

from arXiv · show

This paper presents a comparison of six machine learning (ML) algorithms: GRU-SVM (Agarap, 2017), Linear Regression, Multilayer Perceptron (MLP), Nearest Neighbor (NN) search, Softmax Regression, and Support Vector Machine (SVM) on the Wisconsin Diagnostic Breast Cancer (WDBC) dataset (Wolberg, Street, & Mangasarian, 1992) by measuring their classification test accuracy and their sensitivity and specificity values. The said dataset consists of features which were computed from digitized images of FNA tests on a breast mass (Wolberg, Street, & Mangasarian, 1992). For the implementation of the ML algorithms, the dataset was partitioned in the following fashion: 70% for training phase, and 30% for the testing phase. The hyper-parameters used for all the classifiers were manually assigned. Results show that all the presented ML algorithms performed well (all exceeded 90% test accuracy) on the classification task. The MLP algorithm stands out among the implemented algorithms with a test accuracy of ~99.04%.

1 INTRODUCTION

Breast cancer is a major research concern, and prior work has applied machine learning to classify tumors using the WDBC dataset. This paper extends that line of study by introducing GRU-SVM alongside other classifiers.

  • Breast cancer is among the most common cancers and represents 15% of all new cancer cases in the United States.Source: §1 · p. 1
  • Machine learning may assist medical practitioners’ decision-making and support research using the growing volume of breast-cancer data.Source: §1 · p. 1
  • Prior studies applied machine-learning algorithms to breast-cancer classification with the WDBC dataset and reported significant results.Source: §1 · p. 1
  • This paper introduces GRU-SVM and applies it with other machine-learning algorithms to breast-cancer classification using WDBC.GRU-SVM combines a gated recurrent unit with a support vector machine.Source: §1 · p. 1

2.1 Machine Intelligence Library

The study implemented its machine-learning algorithms using Google TensorFlow and supporting scientific-computing libraries.

  • Google TensorFlow implemented the machine-learning algorithms, supported by matplotlib, numpy, and scikit-learn.Source: §2.1 · p. 1

2.2 The Dataset

The WDBC dataset represents breast-mass FNA images through measurements describing cell-nucleus characteristics. It contains 569 data points and 30 derived features.

  • WDBC features were computed from digitized fine-needle-aspirate images of breast masses and describe cell-nucleus characteristics.Source: §2.2 · p. 1
  • Figure 1 shows digitized FNA images labeled benign and malignant.Source: Figure 1 · p. 1
  • The dataset contains 569 data points: 212 malignant and 357 benign.Source: §2.2 · p. 1
  • The dataset includes radius, texture, perimeter, area, smoothness, compactness, concavity, concave points, symmetry, and fractal dimension.Source: §2.2 · p. 2

2.3 Dataset Preprocessing

Preprocessing standardized the dataset to avoid inappropriate relevance assignment, used thresholded linear regression as a classifier, and measured model loss with mean squared error.

  • The dataset was standardized before modeling to avoid inappropriate assignment of feature relevance.Source: §2.3 · p. 2
  • Linear regression was converted into a classifier by thresholding its output.Source: §2.4.2 · p. 2
  • Mean squared error was used to measure model loss.Source: §2.4.2 · p. 2
  • Standardization uses each feature’s mean and standard deviation and was implemented with scikit-learn’s StandardScaler().fit_transform().Source: §2.4.2 · p. 2

2.4 Machine Learning (ML) Algorithms

The paper applies several ML algorithms to breast cancer classification, including a GRU-SVM architecture that feeds GRU representations into an L2-SVM predictor. The section also describes the classifiers’ optimization procedures, distance measures, and decision outputs.

  • The study implements ML algorithms including GRU-SVM, Nearest Neighbor search, Softmax Regression, SVM, MLP, and Linear Regression.Source: §2.4 · p. 2, §2.4.1 · p. 2

GRU-SVM

  • GRU-SVM uses the GRU cell state h_t as the L2-SVM predictor input instead of a conventional Softmax classifier.Source: §2.4.1 · p. 2
  • The GRU-SVM learning parameter is optimized with L2-SVM loss and Adam, producing scores mapped to benign (-1) or malignant (+1) diagnoses.Source: §2.4.1 · p. 2

Nearest Neighbor

  • Nearest Neighbor search identifies the closest point using both L1 Manhattan and L2 Euclidean distances.Source: §2.4.4 · p. 2
  • Softmax Regression produces class probabilities, whereas Linear Regression produces raw class scores that require thresholding for classification.Source: §2.4.5 · p. 3

Support Vector Machine

  • SVM classification uses a separating hyperplane, with C controlling the penalty and L2-SVM providing differentiable, more stable results than L1-SVM.Source: §2.4.6 · p. 3

2.5 Data Analysis

The experiment separates training and testing into two phases using a 70%/30% dataset partition and evaluates classification with accuracy, error rates, and true-positive and true-negative rates.

  • 70% of the dataset was used for training and 30% for testing.Source: §2.5 · p. 3
  • The experiment had separate training and test phases.Source: §2.5 · p. 3
  • Evaluation parameters included Test Accuracy, Epochs, data-point count, FPR, FNR, TPR, and TNR.Source: §2.5 · p. 3

3 RESULTS AND DISCUSSION

The experiments compare ML classifiers on WDBC using a 70%/30% training/testing split, with results summarized through accuracy and visualized feature separability. All reported algorithms performed strongly, while MLP and L2-SVM achieved particularly high accuracy; comparisons with prior SVM results are constrained by different data partitions.

Test comparison

  • 96.09% test accuracy for L2-SVM exceeded the 89.28% reported for Gaussian-RBF SVM in the prior study.The current study used 70% training data, compared with 60% in the prior study.Source: §3 · p. 3

Training results

  • 97.734375% average training accuracy was achieved by L2-SVM, followed by Softmax Regression at 97.366573% and MLP at 96.9286785%.Linear Regression reached 92.8906257%, while GRU-SVM reached 90.6857639%; Nearest Neighbor had no recorded training accuracy because it requires no training.Source: §3 · p. 4

Model discussion

  • GRU-SVM achieved 93.75% test accuracy, with its mid-level performance hypothetically attributed to GRU nonlinearities and sensitivity to arbitrary weight initialization.The paper also states that arbitrary initialization may limit result reproducibility under identical configurations.Source: §3 · p. 4

Feature visualization

  • The WDBC features appear visually separable by a linear function in scatter plots of mean, error, and worst features.Figures 3–5 show feature groups x0−x9, x10−x19, and x20−x29, respectively.Source: Figure 3 · p. 3, Figure 4 · p. 4, Figure 5 · p. 4

4 CONCLUSION AND RECOMMENDATION

The study concludes that the evaluated machine-learning algorithms performed strongly on binary breast-cancer classification, while recommending cross-validation to strengthen performance estimates and hyper-parameter selection.

Conclusion

  • All presented machine-learning algorithms exhibited high performance in distinguishing benign from malignant tumors.The paper also reports satisfactory statistical measures for the classification problem.Source: §4 · p. 5

Recommendation

  • k-fold cross-validation should be employed to provide a more accurate measure of model prediction performance.The authors also state that it could help determine more optimal hyper-parameters for the algorithms.Source: §4 · p. 5

Related papers

Loading 1711.07831v4…