Attention-guided and MIL-constrained CycleGAN for high-fidelity virtual p16 and Ki-67 staining.

Cervical intraepithelial neoplasia (CIN) grading relies on both morphological assessment from hematoxylin and eosin (H&E) slides and biomarker expression from immunohistochemistry (IHC), such as p16 and Ki-67. However, conventional IHC staining is time-consuming, costly, and requires additional tissue sections. To address these limitations, we propose an attention-guided and multi-instance learning (MIL) constrained CycleGAN framework for high-fidelity virtual IHC staining generation from H&E images. To better utilize the correspondence between H&E and IHC slides, we adopt a weakly paired training strategy. Specifically, whole-slide image registration is first applied to align H&E and corresponding IHC slides, followed by patch-level pairing for model training. Considering the inherent spatial misalignment between adjacent tissue sections, the proposed framework integrates attention-guided constraints and MIL-based supervision to enhance structural consistency and biomarker distribution reliability. In addition to quantitative image-level and diagnostic evaluations, we further conduct an expert-based clinical assessment. Two experienced pathologists independently evaluated CIN grading using H&E slides with and without virtual staining. The results demonstrate that the inclusion of virtual IHC improves diagnostic performance, supporting the clinical utility of the proposed method. To evaluate generalization capability, we additionally validate the model on an external dataset (ACROBAT) under cross-domain conditions. This experiment is designed to assess robustness rather than direct task alignment, providing supplementary evidence of model transferability. We also acknowledge the potential risk of false-positive biomarker synthesis in virtual staining. To address this, we incorporate statistical consistency constraints and provide quantitative evaluation, including diagnostic consistency and expert assessment, which together support the biological plausibility of the generated staining patterns. Overall, the proposed framework provides a practical and clinically relevant solution for virtual IHC generation, achieving superior performance across key evaluation metrics including accuracy (ACC), area under the curve (AUC), and kappa scores, while also demonstrating improved diagnostic support and cross-domain generalization capability.