Overview
Researchers propose a novel deep learning framework designed for the automated grading of Cervical Intraepithelial Neoplasia (CIN) and the prediction of clinical Swede scores. The framework incorporates a dual-stream cross-attention architecture that simulates expert colposcopic visual reasoning by fusing paired multimodal cervigrams. This approach specifically evaluates comparative tissue responses. A custom composite loss function was developed to address severe class imbalances and inconsistencies in scoring across the five components of the Swede score. The development also included the introduction of the BUET Multi-Center Colposcopy Dataset, which is a new, multi-center cohort specifically designed and annotated for both Swede score prediction and CIN grading.
Research Context
Cervical cancer represents a significant global health challenge. Its disease burden is disproportionately high in low- and middle-income countries (LMICs). This disparity is attributed to a shortage of trained specialists and the subjective nature inherent in colposcopy-based screening methods. The proposed deep learning framework directly addresses these identified challenges by aiming to provide AI-assisted colposcopy screening tools. The objective is to support risk-based triage in healthcare settings with limited resources.
Approach
The research introduces a dual-stream cross-attention architecture. This architecture's design mimics the visual reasoning process employed by an expert colposcopist. It functions by explicitly fusing paired multimodal cervigrams. The purpose of this fusion is to evaluate comparative tissue responses, which is a critical aspect of colposcopic assessment. To enhance performance and address specific data characteristics, a custom composite loss function was integrated into the framework. This function was engineered to mitigate the impact of severe class imbalances and inconsistencies in scoring that were observed across the five individual components of the Swede score.
The development and evaluation of this framework leveraged a new dataset: the BUET Multi-Center Colposcopy Dataset. This dataset is a novel, multi-center cohort. It was designed and annotated specifically to facilitate research in Swede score prediction and CIN grading within a machine learning context.
Findings
- For three-class CIN grading, the proposed framework achieved an accuracy of 71.85%.
- The framework demonstrated an 86.23% AUC-ROC for three-class CIN grading.
- This performance in CIN grading was indicated to outperform existing methods.
- For the prediction of individual Swede score components, the architecture yielded AUC-ROC values ranging from 75.7% to 88.4%.
- The custom composite loss function consistently produced improvements in F1-scores across the Swede score components.
- The total predicted Swede Score, which is scaled between 0 and 10, exhibited a Mean Absolute Error (MAE) of 1.489.
- The dataset and source code for the proposed framework are publicly available.
Why This Matters
The results indicate that the developed method can contribute to the creation of AI-assisted colposcopy screening tools. Such tools are positioned to support risk-based triage strategies within resource-limited healthcare settings. This addresses the challenge of cervical cancer's disproportionate burden in LMICs, where specialist shortages and subjective screening methods currently prevail.