Generalizable Cervical Cancer Screening via Large-scale Pretraining and Test-Time Adaptation

Recently, the HKUST team, in collaboration with several leading medical centers, including PKU Shenzhen Hospital, Prince of Wales Hospital, and The First Affiliated Hospital of Sun Yat-sen University, has completed a breakthrough project on cervical cancer screening, leveraging large-scale pretraining and test-time adaptation to achieve robust generalization across diverse clinical settings. This work introduces Smart-CCS—a new paradigm designed to overcome longstanding challenges in cytology-based cancer screening.

Existing Challenges

In recent years, although AI-assisted cytology has demonstrated promising results, current screening systems are plagued by several critical issues:

Figure 1: Key challenges in clinical cytology highlighting cell morphology complexity and data variability.

Cytomorphorlogy Similarity: Cervical cell images often exhibit subtle differences among various cell types and conditions. The overlapping morphological features and the sparse distribution of lesion cells in samples make accurate detection difficult.

Sparsity of abnormal cells: The rare occurrence of lesion cells in cervical specimens poses a significant challenge for traditional AI models, which are often trained on limited datasets with imbalanced class distributions.

Gigapixel Image Analysis: The high-resolution nature of cytology slides results in large image sizes, making it computationally expensive to process and analyze these images in real time.

Limited Generalizability: Many existing models are tailored for specific datasets or conditions, causing significant performance drops when deployed in diverse clinical environments where variations are common.

Our Work

Data: Addressing these challenges, the research team assembled an extensive dataset named CCS-127K, encompassing cytology slides from 48 medical institutes. The dataset contains over 127,000 whole-slide images, annotated with detailed cell-level and slide-level labels, providing a rich source of information for training and evaluation.

Figure 2: Overview of the CCS-127K dataset, a comprehensive collection of cytology slides from diverse clinical settings. a. Classwise distribution of 127,471 WSIs collected from 48 medical centers, including 124118 samples from 45 retrospective centers and 3,353 samples from 3 prospective centers. b. Abnormal cell annotation statistics: 104,979 abnormal lesion cells were annotated into 6 categories, ASC-US, LSIL, ASC-H, HSIL, SCC, and AGC, termed as CCS-Cell dataset. c. Designed flowchart of the study for the development and validation of the proposed Smart-CCS system. The orange flow represents retrospective studies, while the blue flow represents prospective studies.

A study flow chart is shown in Figure 2c, illustrating the development and validation of the Smart-CCS system through retrospective and prospective studies. The orange flow represents retrospective studies, while the blue flow represents prospective studies. In the retrospective studies, the model was trained and evaluated on the CCS-127K dataset, while in the prospective studies, the model was tested in real clinical settings to assess its generalizability.

Method: The novel screening framework, Smart-CCS, is built upon a three-stage process:

Large-scale Self-Supervised Pretraining: By leveraging the expansive CCS-127K dataset, the model first learns general cytological representations without requiring labor-intensive annotations.

Model Finetuning for Cytology Screening: Using both slide-level and cell-level supervision, the model is fine-tuned to specialize in identifying abnormal cell clusters within high-resolution cervical specimens.

Test-Time Adaptation: In a key innovation, the model dynamically updates its weights during inference. This test-time adaptation allows the system to adjust to unforeseen variations in specimen quality and staining, ensuring robust performance across different clinical settings.

Figure 3: The Smart-CCS paradigm consists of three sequential stages. a. the pretraining stage, which involves large-scale self-supervised pretraining on diverse cytology images from various centers to build a generalizable feature extraction model. b. the finetuning stage, which specializes the pretrained model for cancer screening tasks, including two components: an abnormal cell detector for identifying abnormal cells and a WSI classifier for slide-level predictions. c. the adaptation stage, which further optimizes trained model for diverse clinical settings via adapting and refining predictions.

Results

Smart-CCS redefines the standards for cervical cancer screening with striking performance improvements demonstrated across both retrospective and prospective studies:

Retrospective Evaluations: On 11 internal test datasets, Smart-CCS achieved an impressive AUC of 0.965 and a sensitivity of 0.913, indicating high diagnostic accuracy, especially in extremely imbalanced class distributions. On 6 external test datasets, the model maintained an AUC of 0.950, showcasing its robustness and generalizability.

Figure 4: Overview of Smart-CCS performance across multiple retrospective evaluations. Internal and external data distribution, along with the results of cervical cancer screening evaluations are exhibited in the figure.

We also conduct ablation studies to demonstrate the effectiveness of the proposed pretraining and test-time adaptation on the external test datasets. From the results, we observed when both pretraining and adaptation were applied, Smart-CCS achieved a 6.3% AUC improvement against the CCS baseline model. The demonstrated effectiveness of the Smart-CCS paradigm in external testing underscores its potential clinical applicability in complex scenarios.

Figure 5: Ablation studies on the external test dataset. We evaluated the proposed model with different settings, Base denotes the typical two-step CCS model, w/ P is introducing pretraining, w/ P&A refers to our proposed Smart-CCS with pretraining and adaptation.

Prospective Evaluations: In live clinical settings across three prospective centers, the model reached AUCs of 0.947, 0.924, and 0.986, respectively, confirming its adaptability and generalization. Furthermore, histological examination results confirmed the model’s high sensitivity and specificity in cervical cancer detection, reinforcing its reliability in clinical applications. The author team also generated detailed visualizations of slide- and cell-level predictions, which revealed that the screening decisions made by Smart-CCS closely mirror to clinical assessments, adding a layer of transparency to its outputs.

Figure 6: Overview of Smart-CCS performance across multiple prospective evaluations. a. Grade distribution among three prospective centers (PC1, PC2 and PC3) and reported evaluation results from our Smart-CCS system. b. The diagnostic performance of Smart-CCS in cancer detection using histological results as the gold standard. c. Visualizations from cervical cancer screenings, which include interpretable results at both the cell-level and slide-level, derived from three sample tests.

Conclusion

Smart-CCS marks a significant advancement in AI-powered cervical cancer screening by addressing the inherent variability in clinical data and adapting in real time to new conditions. The study emphasizes the importance of large-scale pretraining and dynamic test-time adaptation, recommending further exploration into sophisticated computational cytology models and broader validation across diverse clinical settings.

For more detailed results and technical insights, please refer to the original paper on arXiv.

If you have any questions or suggestions, please feel free to contact the authors. For bibtex citation, please use the following format:

@article{jiang2025smartccs,
      title={Generalizable Cervical Cancer Screening via Large-scale Pretraining and Test-Time Adaptation},
      author={Jiang, Hao and Jin, Cheng and Lin, Huangjing and Zhou, Yanning and Wang, Xi and Ma, Jiabo and Ding, Li and Hou, Jun and Liu, Runsheng and Chai, Zhizhong and Luo, Luyang and Shi, Huijuan and Qian, Yinling and Wang, Qiong and Li, Changzhong and Han, Anjia and Chan, Ronald Cheong Kin and Chen, Hao}
      journal={arXiv preprint arXiv: 2502.09662},
      year={2025}
}