(W0930-05-30) Machine-Learning-Based Assessment of Breast Tumor Malignancy Through Nuclear Morphological Features

Purpose: Approximately 30% of all new female cancers diagnosed annually are breast cancers, making it one of the most common and lethal diseases among women worldwide. To enhance prognosis and improve treatment outcomes, breast cancer must be diagnosed early and accurately. Although the diagnosis of breast cancer has been greatly improved through the use of routine diagnostic techniques—such as mammography, MRI, PET, and CT scans—these techniques are not without their shortcomings. Consequently, we were motivated to improve the detection of breast cancer by utilizing statistical testing and a variety of machine learning algorithms.

Methods: We used a publicly available dataset comprising more than 500 patient breast tumor nuclei (Breast Cancer Wisconsin Diagnostic Database). Each nucleus had quantitative morphological attributes: area, perimeter, radius, texture, smoothness, compactness, concavity, concave points, symmetry, and fractal dimension. The dataset was preprocessed prior to further analysis. Mann-Whitney non-parametric testing was used to assess differences between malignant and benign tumor groups. The ability of each feature—or any combination thereof—to separate benign from malignant tumors was further analyzed using unsupervised k-means clustering. The quality of cluster separation of all possible feature pairs was visually evaluated in two-dimensional scatter plots. Supervised machine learning classifiers were evaluated next. The data set was randomly divided into training (80%, 456 instances) and testing (20%, 113 instances) sets. Ten-fold cross-validation was conducted to tune the performance of seven machine learning classifiers. The performance of the classifiers was compared based on five performance metrics—area under the ROC curve (AUC), accuracy, F1-score, precision, and recall. SHAP analysis was utilized in order to determine the most important features accountable for model predictions.

Results: Nine out of the ten nuclear morphometric characteristics—excluding fractal dimension—differed significantly between malignant and benign tumors. The area, perimeter, and concavity characteristics had larger median values and more dispersed value distributions in malignant tumors, suggesting that these characteristics are indicative of structural transformations associated with malignancy. Certain features were better at discriminating between tumor types based on unsupervised clustering. In particular, minimally overlapping clusters were consistently produced by combinations of area, perimeter, radius, concavity, and concave points. Scatter plots based on fractal dimension and texture, however, resulted in poor cluster separation, suggesting limited discriminatory value. All of the seven supervised classifiers trained well, as all of the metrics were more than 89%. The neural network classifier was better than the others, with the highest metrics in both the training and test sets. The neural network achieved accuracy, precision, recall, and F1-score of 96.3%, and AUC of 98.8% on the training set. It accurately identified 150 of 159 cancerous cases and 289 of 297 non-cancerous cases. In the test set, it misclassified one benign and five malignancies and achieved 98.2% AUC, accuracy and recall of 94.7%, precision of 94.9%, and an F1-score of 94.7%. The overall performance metrics and confusion matrices of the supervised classifiers, i.e., the neural network and gradient boosting models, are combined in Figure 1. SHAP analysis provided deeper insight into the decision-making processes of each classifier. Concave points were the most significant feature in four classifiers (support vector machine, neural network, gradient boosting, and random forest) and were consistently ranked among the most important features across all models. Area was the most significant feature in the remaining three classifiers. Concavity was also consistently ranked among the top three significant features. The feature importance rankings according to SHAP analysis are plotted in Figure 2. Conversely, in three classifiers, fractal dimension was the least significant feature, and in the others, it ranked near the bottom. Features such as symmetry, smoothness, and compactness had relatively little effect on classifier predictions.

Conclusion: Through a combination of statistical and machine learning analyses, we demonstrated that specific nuclear morphological features—particularly concave points, area, and concavity—have a strong association with breast tumor malignancy and may potentially serve as morphological markers to improve the detection of breast cancer.


Figure 1: Classification performance of the supervised classifiers against the testing dataset. (a) Table summarizing the quantitative metrics of all supervised classifiers. The classifier order is based on the values of accuracy, F1, precision, and recall. (b, c) Confusion matrices of the two best performing classifiers: (b) neural network and (c) gradient boosting.


Figure 2: Influence of different breast tumor nuclear morphological features on the classifier outputs. (a, b) SHAP values and ranking of features considered by the two best performing classifiers: (a) neural network and (b) gradient boosting.