Uncertainty-aware adaptive fusion with graceful degradation for trustworthy cervical cytology
TL;DR
An uncertainty-aware deep learning framework that integrates an ensemble image encoder comprising ResNet50, DenseNet121, and EfficientNet-B0 with dedicated clinical and genomic encoders through an adaptive fusion mechanism that serves as a proof-of-concept for uncertainty-guided multimodal fusion.
Abstract
Trustworthy cervical cancer screening requires not only accurate classification but also reliable estimation of predictive uncertainty to identify cases that should be referred to clinical experts. This study presents an uncertainty-aware deep learning framework that integrates an ensemble image encoder comprising ResNet50, DenseNet121, and EfficientNet-B0 with dedicated clinical and genomic encoders through an adaptive fusion mechanism. Modality-specific predictions are subsequently combined using a convex weighting scheme driven by predictive entropy estimated via Monte Carlo Dropout. As paired patient-level multimodal data were unavailable, clinical features were derived from an independent risk-factor cohort and genomic features were synthetically generated. The proposed framework serves as a proof-of-concept for uncertainty-guided multimodal fusion. When auxiliary modalities are uninformative, the framework automatically suppresses their influence, maintaining identical test-set predictions across all image-containing configurations, while a static equal-weight baseline reduced the AUC from 0.993 to 0.982. Evaluated on a combined SIPaKMeD and Herlev benchmark of 6,849 cervical-cell images, the framework attains an AUC-ROC of 0.993 (95% CI 0.988–0.997), 97.08% accuracy, 98.07% sensitivity, and 96.08% specificity, with well-calibrated probabilities (Brier 0.025; ECE 0.014). Cross-dataset transfer experiments revealed a notable generalisation gap, highlighting the need for external validation, while confidence-based selective prediction demonstrated the framework’s potential for human-in-the-loop triage.