Unsupervised Anomaly Detection for Cervical Cancer Screening on Unpaired Public Data: A Leakage-Aware Empirical Study

Cervical cancer screening draws on cytology, colposcopy and clinical history, yet public datasets are organized by modality and almost never share patients, so patient-level multimodal models cannot be trained or validated on them. We therefore ask a narrower question that such data can answer: how well do unsupervised normative anomaly detectors work on each modality under careful evaluation? We built three independent pipelines (cytology images, colposcopy images and a clinical risk-factor table) using frozen or distilled visual embeddings and five scorers: an autoencoder, Isolation Forest, One-Class SVM, Mahalanobis distance and k-nearest-neighbor (kNN) distance. Evaluation used group-level splits, a contamination study, cross-dataset tests and clustered bootstrap intervals. On cytology, plain kNN on frozen features reached a macro area under the receiver operating characteristic curve (AUROC) of 0.93–0.95, whereas our equal-weight ensemble reached only 0.65–0.78, and distillation gave no benefit. On colposcopy, every scorer was at chance (AUROC 0.44–0.51). Clinical history gave a weak signal (AUROC 0.64). Embeddings identified the source dataset with 94–99% accuracy, so cytology scores may partly reflect acquisition differences rather than pathology. We claim no diagnostic validity and no patient-level fusion.