A Stain-Invariant and Calibrated Deep Learning Framework for Automated Cervical Cytology Classification

Pap-smear cytology, a method for the screening of cervical cancer, is exposed to a number of issues such as stain variability, morphological alterations, class imbalance, and unavailability of dependable confidence estimation in automated systems. To overcome these problems, the article comes up with a stain-invariant and well-calibrated multi-scale deep learning system named SiCal-CytoNet for cervical cell classification. It not only supports stain separation, domain generalization, adaptive multi-scale feature fusion, prototype-guided classification and probability calibration to improve diagnosis accuracy but also ensures reliability. The SiCal-CytoNet achieved a classification accuracy of 98.9%, macro-F1 score of 98.7%, AUROC of 99.7%, and AUPRC of 99.4% on the SIPaKMeD dataset. It was, at the same time, well-calibrated as shown by the low Brier score of 0.012 and Expected Calibration Error (ECE) of 1.7%. On top of that, the model was able to handle stain illumination blurring, and variations in styles effectively. The introduction of a confidence-based selective prediction process enabled the model to individually manage 88.1% of the cases while ensuring high diagnostic reliability. The findings very much indicate that the triple combination of stain-invariant learning, multi-scale feature representation, and calibrated prediction results in an accurate, robust, and clinically reliable solution for an AI-empowered cervical cancer screening and cytology decision support.