LEAKAGE AND EVALUATION BIAS IN MACHINE LEARNING FOR BREAST CYTOLOGY DIAGNOSIS: A REVIEW
Keywords:
Breast Cancer Diagnosis, Fine-Needle Aspiration, Wisconsin Diagnostic Breast Cancer Data, Machine Learning, Feature Selection, Stacked Generalization, Cross-Validation, Selection BiasAbstract
Breast cancer remains the most frequently diagnosed cancer among women, and fine-needle aspiration (FNA) cytology offers a minimally invasive route to separating benign from malignant masses. Over several decades, machine-learning classifiers trained on nuclear features taken from digitised cytology images have reported very high accuracy on public benchmarks, above all the Wisconsin Diagnostic Breast Cancer (WDBC) data set. This paper reviews the literature on benchmark data, classification algorithms, feature selection, ensemble learning and evaluation methodology, and compares the surveyed studies in terms of scope, strengths and limitations. The comparison shows that headline accuracies are hard to compare, because studies differ in how data are split, in whether pre-processing and feature selection were fitted inside or outside the validation loop, and in the metrics that are reported. Further problems are the reliance on a single split, little attention to false negatives and decision thresholds, a weak separation between the effect of feature selection and the effect of the classifier, and uncertainty about whether stacked ensembles help on small clinical data sets. Based on these problems, the paper outlines a future direction in which every fitted step is learned only from training data and models are judged by repeated resampling and operating-point analysis.