Hybrid feature fusion in cervical cancer cytology: a novel dual-module approach framework for lesion detection and
Shurong Niu1, Lili Zhang1, Lina Wang1
1Department of Obstetrics and Gynecology, Shanxi Medical University Second Hospital, Taiyuan, China.
Objective:
Cervical cancer screening through cytology remains the gold standard for early detection, but manual analysis is time-consuming, labor-intensive, and prone to inter-observer variability. This study proposes an automated deep learning-based framework that integrates lesion detection, feature extraction, and classification to enhance the accuracy and efficiency of cytological diagnosis.
Materials And Methods:
A dataset of 4,236 cervical cytology samples was collected from six medical centers, with lesion annotations categorized into six diagnostic classes (NILM, ASC-US, ASC-H, LSIL, HSIL, SCC). Four deep learning models, Swin Transformer, YOLOv11, Faster R-CNN, and DETR (DEtection TRansformer), were employed for lesion detection, and their performance was compared using mAP, IoU, precision, recall, and F1-score. From detected lesion regions, radiomics features (n=71) and deep learning features (n=1,792) extracted from EfficientNet were analyzed. Dimensionality reduction techniques (PCA, LASSO, ANOVA, MI, t-SNE) were applied to optimize feature selection before classification using XGBoost, Random Forest, CatBoost, TabNet, and TabTransformer. Additionally, an end-to-end classification model using EfficientNet was evaluated. The framework was validated using internal cross-validation and external testing on APCData (3,619 samples).
Results:
The Swin Transformer achieved the highest lesion detection accuracy (mAP: 0.94 external), outperforming YOLOv11, Faster R-CNN, and DETR. Combining radiomics and deep features with TabTransformer yielded superior classification (test accuracy: 94.6%, AUC: 95.9%, recall: 94.1%), exceeding both single-modality and end-to-end models. Ablation studies confirmed the importance of both the detection module and hybrid feature fusion. External validation demonstrated high generalizability (accuracy: 92.8%, AUC: 95.1%). Comprehensive statistical analyses, including bootstrapped confidence intervals and Delong's test, further substantiated the robustness and reliability of the proposed framework.
Conclusions:
The proposed AI-driven cytology analysis framework offers superior lesion detection, feature fusion-based classification, and robust generalizability, providing a scalable solution for automated cervical cancer screening. Future efforts should focus on explainable AI (XAI), real-time deployment, and larger-scale validation to facilitate clinical integration.
