Abstract
The thinprep cytologic test is widely used for cervical cancer diagnosis, with results reliant on cytotechnicians screening specific cell categories on slides, lacking robustness. This paper introduces a two-stage quantitative detection framework for whole slide cervical images, aiding pathologists in effectively assigning lesion grades. Our approach utilizes a You Only Look Once network with an attention module and multi-scale feature fusion to enhance representation refinement and improve cervical cell classification precision (0.8647) and true positive rate (95.8%). We incorporate a quantitative DNA description, leveraging the Matthew effect to refine diagnostic contributions, establishing a clearer standard for cell proliferation assessment. By extracting time series features and leveraging global smear information, our model enhances detection robustness, enriches the screening system, and resists false cell classification influences (smear-level accuracy, sensitivity, and specificity of 0.9193, 0.9285, 0.9234). We elucidate the significance of time series features in cervical cancer detection, demonstrating efficient global smear information utilization. Evaluation on clinical datasets underscores the relationship between time series features and patient physiological states, facilitating seamless integration into current diagnostic systems, enriching detection principles, and achieving comparable grading accuracy to professional pathologists in cervical intraepithelial neoplasia assessment.