Ensemble Learning to Categorize the Bethesda System of Reporting of Cervical Cytology.

Abstract

Aims and objectives: To evaluate the performance of a soft-voting ensemble deep learning (DL) model incorporating five transfer learning (TL) architectures for categorizing cervical smears according to the Bethesda system.

Materials and methods: A total of 259 cervical cytology cases processed by SurePath liquid-based cytology were included, yielding 1016 representative microphotographs. Images were divided into training (70%), validation (15%), and test (15%) sets. An ensemble model comprising VGG16, ResNet50, InceptionV3, DenseNet121, and MobileNetV2 was trained using TL with soft-voting integration. Diagnostic performance was assessed using standard metrics, including 95% confidence intervals (CI).

Results: In the test set, the model correctly classified 27/31 negative for intraepithelial lesions and malignancy (NILM) images and 56/58 carcinoma images. For NILM, the sensitivity was 0.871 (95% CI: 0.72-0.95), specificity 0.977 (95% CI: 0.94-0.99), accuracy 0.956 (95% CI: 0.92-0.98), and area under the curve of receiver operating characteristic (AUROC) 0.98 (95% CI: 0.96-1.00). For carcinoma, the sensitivity was 0.966 (95% CI: 0.88-1.00), specificity 0.960 (95% CI: 0.87-0.99), accuracy 0.962 (95% CI: 0.93-0.99), and AUROC 0.99 (95% CI: 0.98-1.00). Intermediate categories demonstrated variable performance, with lower sensitivity for low-grade squamous intraepithelial lesion (0.20), though the specificity remained high (1.00). Overall, AUROC values were high across all categories (0.93-0.99). Among all the models, the ensemble learning (EL) model had the best performance.

Conclusions: The proposed EL approach demonstrates high diagnostic accuracy and excellent discrimination for clinically critical categories such as high-grade squamous intraepithelial lesion (HSIL) and carcinoma. This model highlights the potential of DL-based ensemble frameworks in improving the reliability of cervical cytology interpretation. Larger datasets and whole-slide image analysis may further refine the performance for intermediate-grade lesions.