Objective:

To develop a robust deep learning framework for cervical cell classification that specifically addresses the challenges of spatial irrelevance and class imbalance in cytopathology.

Key Findings:

• Achieved a classification accuracy of 99.63% on the SiPaKMeddataset, outperforming 15 competitive models, indicating superior performance in cervical cancer classification.
• Demonstrated 98.62% accuracy on an independent Liquid-Based Cytology dataset, confirming robustness and generalizability across different datasets.
• Visual interpretability analyses showed the model focuses on biologically relevant cellular regions, enhancing trust in its predictions.

Interpretation:

DeepInsight-Net effectively mitigates the challenges of spatial irrelevance and class imbalance, making it a promising tool for enhancing the accuracy of cervical cancer screening in clinical settings.

Limitations:

• The study may be limited by the datasets used, which may not fully represent the diversity of broader populations, potentially affecting the generalizability of the findings.
• There is a potential risk of overfitting to specific datasets despite high accuracy, which could limit the model's performance on unseen data.

Conclusion:

DeepInsight-Net presents a reliable computer-aided diagnosis tool for cervical cancer screening, combining advanced feature learning and loss optimization.