Cervical Intraepithelial Neoplasia (CIN1-3) Disease Grading Using a Mixture of Experts Approach.

Accurate grading of cervical intraepithelial neoplasia (CIN1-3) from colposcopic images is clinically critical yet challenging due to subtle inter-grade morphology and substantial imaging variability. We propose an attention-guided mixture-of-experts (MoE) framework that ensembles five pretrained DenseNet-121 experts and employs an attention mechanism over pooled intermediate features to drive a gating network that adaptively weights expert outputs for each image. Operating on feature representations rather than raw pixels allows the gating network to perform input-specific expert selection and improves robustness to ambiguous cases. Using the Intel & MobileODT cervical screening dataset with a strict patient-wise 70/10/20 split, we report mean performance over five runs with 95% confidence intervals. On the independent test set, the proposed MoE achieves 74.0% ± 1.6 accuracy and 72.1% ± 1.8 F1, with per-class AUCs of 0.88 (CIN1), 0.82 (CIN2), and 0.85 (CIN3). The method yields statistically significant improvements over single-network DenseNet-121 baselines and alternative MoE backbones (MobileNet, EfficientNet, ShuffleNet) (p < 0.01). Ablation studies show that attention-guided gating contributes approximately 5-8% absolute accuracy gain over uniform weighting, and that five experts provide the optimal accuracy-efficiency balance. We further present attention visualizations and limited external validation to assess interpretability and generalizability. Although performance remains below that of recent transformer-ensemble models evaluated on smaller or less diverse test sets, the modular and interpretable MoE architecture offers a practical foundation for integrating segmentation or transformer-based experts to advance clinical utility. Code and trained models will be released to support reproducibility.