A reproducible open-data framework for interpretable breast cancer diagnosis and cytopathology education
- 1Department of Oncology, The First Affiliated Hospital of Anhui Medical University
- 2Department of Nuclear Medicine, The First Affiliated Hospital of University of Science and Technology of China (USTC)
- 3Department of Oncology, The 901 Hospital of the Joint Logistics Support Force of the People’s Liberation Army of China
Abstract
Introduction: Quantitative nuclear morphology links cytopathological assessment with interpretable diagnostic modeling and provides a practical foundation for teaching medical artificial intelligence. This study evaluated a reproducible framework that connects nuclear feature interpretation, malignant-versus-benign classification, and evidence-based assessment of diagnostic decisions.
Methods: The Wisconsin Diagnostic Breast Cancer dataset included 569 fine-needle aspiration samples (212 malignant and 357 benign) represented by 30 image-derived nuclear features. Seven supervised classifiers were compared using nested stratified cross-validation with five outer folds and four inner folds. Assessment covered ROC-AUC, precision-recall performance, probability calibration, training-derived decision thresholds, exploratory decision-curve analysis, feature importance, and leakage-controlled feature reduction.
Results: The radial basis function support vector machine achieved a ROC-AUC of 0.995, while the interpretable linear support vector machine (SVM) achieved 0.994, without a significant pairwise difference from the radial basis function SVM or logistic regression. For the linear SVM, fold-specific Youden thresholds yielded held-out accuracy of 0.981, sensitivity of 0.962, specificity of 0.992, F1-score of 0.974, and Matthews correlation coefficient of 0.959. Sensitivity-prioritized thresholding increased sensitivity to 0.991, with specificity of 0.866. The Brier score was 0.021. Nuclear texture, size, symmetry, and contour-related characteristics contributed prominently to model interpretation. A leakage-controlled 20-feature model retained a ROC-AUC of 0.996 and a precision-recall area under the curve of 0.994.
Discussion: The framework integrates quantitative cytopathology with reproducible model appraisal in a coherent teaching sequence. Nuclear feature distributions and importance patterns connect cellular morphology with diagnostic reasoning; classifier comparisons, calibration, and confusion matrices illustrate distinct dimensions of model performance; and alternative operating thresholds show how predictive probabilities translate into diagnostic decisions. Nested validation and reduced-feature analysis furnish practical examples for teaching data partitioning, feature redundancy, interpretability, and critical evaluation of medical AI. This shared analytical pathway provides case-based material for cytopathology, oncology, and interdisciplinary medical data science education.
Related
- RMSD: an interpretable framework for streaming-compatible multi-source data fusion and skill-gap diagnosis in vocational education
- A federated attention-based stacked LSTM framework for interpretable malaria diagnosis under simulated non-IID federated conditions
- Interpretable attention-deficit hyperactivity disorder diagnosis from functional magnetic resonance imaging via hybrid convolutional neural network–transformer learning and multi-level graph-based comprehensibility
- JICS Education: Death in intensive care. Pathophysiology and diagnosis
- Stress-Testing Meta-Analyses: A Reproducible Python Framework for Robustness Assessment of Continuous Outcomes
BibTeX
@article{li2026reproducible,
title = {{A reproducible open-data framework for interpretable breast cancer diagnosis and cytopathology education}},
author = {Min Li and Xinxin Zhang and Xiaosu Xie and Yu Shi and Xueyang Hu and Nana Huang and Zilu Wang},
journal = {Frontiers in Oncology},
year = {2026},
month = oct,
doi = {10.3389/fonc.2026.1921082},
url = {https://www.frontiersin.org/articles/10.3389/fonc.2026.1921082/full},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - A reproducible open-data framework for interpretable breast cancer diagnosis and cytopathology education
AU - Min Li
AU - Xinxin Zhang
AU - Xiaosu Xie
AU - Yu Shi
AU - Xueyang Hu
AU - Nana Huang
AU - Zilu Wang
PY - 2026
DA - 2026/10/26
JO - Frontiers in Oncology
DO - 10.3389/fonc.2026.1921082
UR - https://www.frontiersin.org/articles/10.3389/fonc.2026.1921082/full
AB - Introduction: Quantitative nuclear morphology links cytopathological assessment with interpretable diagnostic modeling and provides a practical foundation for teaching medical artificial intelligence. This study evaluated a reproducible framework that connects nuclear feature interpretation, malignant-versus-benign classification, and evidence-based assessment of diagnostic decisions. Methods: The Wisconsin Diagnostic Breast Cancer dataset included 569 fine-needle aspiration samples (212 malignant and 357 benign) represented by 30 image-derived nuclear features. Seven supervised classifiers were compared using nested stratified cross-validation with five outer folds and four inner folds. Assessment covered ROC-AUC, precision-recall performance, probability calibration, training-derived decision thresholds, exploratory decision-curve analysis, feature importance, and leakage-controlled feature reduction. Results: The radial basis function support vector machine achieved a ROC-AUC of 0.995, while the interpretable linear support vector machine (SVM) achieved 0.994, without a significant pairwise difference from the radial basis function SVM or logistic regression. For the linear SVM, fold-specific Youden thresholds yielded held-out accuracy of 0.981, sensitivity of 0.962, specificity of 0.992, F1-score of 0.974, and Matthews correlation coefficient of 0.959. Sensitivity-prioritized thresholding increased sensitivity to 0.991, with specificity of 0.866. The Brier score was 0.021. Nuclear texture, size, symmetry, and contour-related characteristics contributed prominently to model interpretation. A leakage-controlled 20-feature model retained a ROC-AUC of 0.996 and a precision-recall area under the curve of 0.994. Discussion: The framework integrates quantitative cytopathology with reproducible model appraisal in a coherent teaching sequence. Nuclear feature distributions and importance patterns connect cellular morphology with diagnostic reasoning; classifier comparisons, calibration, and confusion matrices illustrate distinct dimensions of model performance; and alternative operating thresholds show how predictive probabilities translate into diagnostic decisions. Nested validation and reduced-feature analysis furnish practical examples for teaching data partitioning, feature redundancy, interpretability, and critical evaluation of medical AI. This shared analytical pathway provides case-based material for cytopathology, oncology, and interdisciplinary medical data science education.
SN - 2234-943X
ER -