CALIBRATED RECOGNITION OF MALIGNANT NUCLEAR MORPHOLOGY FROM IMAGE-DERIVED FINE-NEEDLE ASPIRATE FEATURES

Abstract

Image-recognition studies in medicine frequently report high accuracy without separating model selection from performance estimation. This study evaluates whether compact machine-learning models can recognize malignant nuclear morphology from measurements extracted from digitized fine-needle aspirate images while controlling optimism. The public Wisconsin Diagnostic Breast Cancer dataset (569 observations; 212 malignant and 357 benign) was analysed using logistic regression, radial-basis-function support vector machines (RBF-SVM), and random forests. Hyperparameters were selected within four-fold inner cross-validation and evaluated by five-fold outer cross-validation repeated three times. Out-of-fold probabilities were averaged per observation. The RBF-SVM achieved an ROC area under the curve (AUC) of 0.995 (95% bootstrap confidence interval: 0.989–0.999), accuracy of 97.9%, sensitivity of 96.7%, specificity of 98.6%, and Brier score of 0.018. Logistic regression was numerically close, whereas random forest performance was lower. A feature-family ablation showed that the ten “worst” nuclear measurements retained an AUC of 0.992, while standard-error features alone fell to 0.942. These results support morphology-based computer assistance, but not autonomous diagnosis: the single-centre, image-derived feature set lacks external, temporal, and scanner validation.