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

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Pathology

Background:

• Accurate grading of cervical intraepithelial neoplasia (CIN) is crucial for patient management but challenging due to subtle morphological variations and imaging inconsistencies.
• Existing automated methods often struggle with the inherent variability in colposcopic images, necessitating more robust diagnostic tools.

Purpose of the Study:

• To develop and evaluate an attention-guided mixture-of-experts (MoE) framework for accurate grading of CIN1-3 from colposcopic images.
• To improve diagnostic accuracy and robustness in cervical cancer screening by adaptively weighting multiple expert models.

Main Methods:

• Proposed an attention-guided MoE framework ensembling five pretrained DenseNet-121 models.
• Employed an attention mechanism over intermediate features to guide a gating network for adaptive expert weighting.
• Utilized the Intel & MobileODT cervical screening dataset with a patient-wise data split for rigorous evaluation.

Main Results:

• The MoE framework achieved 74.0% accuracy and 72.1% F1 score on the independent test set.
• Demonstrated superior performance over single DenseNet-121 baselines and other MoE backbones (p < 0.01).
• Attention-guided gating provided a significant accuracy gain (5-8%) compared to uniform weighting, with five experts offering optimal balance.

Conclusions:

• The attention-guided MoE framework offers a significant advancement in automated CIN grading from colposcopic images.
• The modular and interpretable architecture provides a practical foundation for integrating diverse AI experts to enhance clinical utility in cervical screening.
• While current performance is below state-of-the-art transformer models, the proposed method shows promise for reproducible and generalizable clinical applications.