Abstract
Early detection of cervical dysplasia is critical for effective treatment, yet automated visual inspection remains challenging due to limited high-quality datasets and domain discrepancies between clinical studies. This paper presents a novel framework for cross-domain knowledge transfer that leverages auxiliary datasets while filtering out inconsistent labels. By utilizing target-domain prototypes to validate auxiliary data and aligning representations at both domain and class levels, the proposed method achieves significant performance gains, including a 4.7% improvement in top-1 accuracy on real-world benchmarks.
Introduction
Cervical cancer remains a significant global health challenge, particularly in low-resource settings where access to advanced diagnostic tools is limited. Visual Inspection with Acetic Acid (VIA) is a cost-effective screening method, but its efficacy is often hampered by the subjectivity of human interpretation. While deep learning offers a path toward automation, the development of robust models is hindered by the scarcity of large-scale, annotated cervical image datasets.
A common strategy to mitigate data scarcity is to leverage data from multiple clinical studies. However, this introduces the "transfer problem," characterized by domain shift (differences in imaging equipment and lighting) and criterion mismatch (subjective differences in diagnostic labeling between clinics). This research addresses these challenges through a prototypical knowledge filtering approach designed to robustly transfer information across disparate cervical image domains.
Methodology
The proposed framework employs a Y-shaped architecture designed to harmonize two distinct datasets into a shared feature space. The architecture is built upon two core modules: Early Domain Alignment (EDA) and Prototypical Semantic Alignment (PSA).
Early Domain Alignment (EDA)
The EDA module focuses on the intermediate feature level to mitigate the effects of domain shift. By employing adversarial training or divergence-based metrics like Multiple Kernel Maximum Mean Discrepancy (MK-MMD), the model is forced to learn features that are invariant to the source of the image. This ensures that the fundamental visual characteristics—such as texture and color changes following acetic acid application—are processed consistently regardless of the specific clinical study from which the image originated.
Prototypical Semantic Alignment (PSA)
The PSA module represents the primary innovation of this work, addressing the issue of "noisy" or inconsistent labels in auxiliary datasets. Instead of assuming all labels in the auxiliary (source) domain are perfectly accurate relative to the target domain, the model establishes "prototypes." These prototypes are the average feature representations for each diagnostic class (e.g., "Normal" vs. "Abnormal") within the target dataset.
The PSA module performs two critical functions:
- Knowledge Filtering: Auxiliary samples are compared against the target prototypes. If an auxiliary sample’s label contradicts its proximity to the target prototypes, it is identified as high-uncertainty. The influence of these samples is then down-weighted during training to prevent the model from learning "incorrect" or inconsistent diagnostic criteria.
- Supervised Contrastive Learning: To further refine the decision boundaries, the model uses supervised contrastive learning to pull representations of the same class closer together in the feature space while pushing different classes apart. This results in a more discriminative and robust classifier.
Experimental Evaluation
The framework was evaluated using 17,002 cervical images across three major clinical studies: the Natural History Study of HPV (NHS), the ASCUS-LSIL Triage Study (ALTS), and a specialized Biopsy Study. Additionally, the Visda-2017 dataset was used to test the generalizability of the method to non-medical contexts.
The results demonstrate that the proposed method (specifically the Ours-adv variant) significantly outperforms existing state-of-the-art models. On the NHS benchmark, the model achieved:
- An absolute improvement of 4.7% in top-1 accuracy.
- A 7.0% increase in precision.
- A 4.6% improvement in F1 score.
- An ROC-AUC of 0.886, indicating high diagnostic reliability.
The performance gains were even more pronounced when the target dataset was extremely small, such as in the Biopsy Study, highlighting the framework's effectiveness in low-resource scenarios.
Discussion and Conclusion
The success of this prototypical approach underscores the importance of treating auxiliary data as a "helpful but potentially unreliable witness." By using target-domain prototypes as anchors, the model can effectively filter out the noise inherent in subjective medical labeling. Furthermore, the dual alignment at both the domain and class levels proves essential for achieving high-performance transfer learning.
While the current model focuses on binary classification of visual images, future research will explore multi-modal integration—combining images with HPV test results and patient history—and more granular classification of cervical intraepithelial neoplasia (CIN) grades. This work provides a robust foundation for more accurate, automated cervical cancer screening tools that can be deployed globally.
Summary updated