Automated Cervical Cancer Screening Via Hybrid Multi-Resolution Image Segmentation and Ensemble Learning: A Comparative Study with SVM and K-Means Approaches

Despite substantial progress in digital pathology, automated cervical cancer screening remains challenging because of overlapping cells, staining variability, and the limited capability of single-resolution segmentation methods to capture morphological variation. This study proposes a hybrid framework that combines multi-resolution image segmentation with stacked ensemble learning to improve screening performance while reducing false-negative diagnoses. Gaussian and Laplacian pyramids are constructed from cervical cytology patches, allowing watershed, active contour, and Gabor-based mean-shift segmentation algorithms to extract complementary structural information at different spatial scales. The resulting features are integrated through a stacked ensemble consisting of a random forest, gradient boosting machine, and multilayer perceptron, with prediction scores combined using a regularized logistic regression meta-learner.

The proposed framework was evaluated on a dataset of 12,500 liquid-based cytology patches annotated according to the Bethesda system and compared with two established baseline approaches: a single-resolution watershed–SVM pipeline and a K-means segmentation–SVM pipeline under identical patient-level five-fold cross-validation. The proposed method achieved a mean classification accuracy of 0.914 ± 0.008, a binary screening AUC of 0.967 ± 0.007, and a sensitivity of 0.891 ± 0.019 for high-grade lesions (HSIL and carcinoma), significantly outperforming both baseline methods (p < 0.001). The false-negative rate for high-grade lesions decreased from 0.238 to 0.109. Ablation analysis further showed that both the multi-resolution segmentation strategy and the stacked ensemble contributed to the overall performance improvements. Inference time remained suitable for routine batch screening, requiring less than five minutes to process a complete slide.

These findings indicate that combining multi-resolution image analysis with ensemble learning provides an effective and computationally practical approach for automated cervical cancer screening, improving diagnostic performance while maintaining efficiency for clinical workflows.