Adaptive Multi-Scale Feature Extraction for Cervical Cancer Classification Using Dynamic Hierarchical Pooling

Abstract

Cervical cancer remains a leading cause of mortality among women worldwide, especially in low-resource settings where access to early screening and treatment is limited. Early detection through accurate and efficient diagnostic methods is critical for improving patient outcomes. This study proposes a novel method for classifying cervical cancer using Dynamic Hierarchical Pooling (DHP). To effectively capture multi-scale characteristics, DHP adaptively modifies the number of pyramid levels and pooling types according to the size of the input image. To be more precise, the module dynamically divides the feature maps into different spatial regions and applies various pooling operations to each region. This adaptive mechanism extracts fine-grained and coarse-grained features crucial for recognizing diverse patterns in cervical pap smear images. To facilitate efficient processing, feature maps are resized to a common size, regardless of the original image size. The Squeeze-and-Excitation (SE) attention module further enhances feature discrimination by dynamically updating attention weights, focusing on the most informative regions of the feature maps. Combining the strengths of Convolutional Neural Networks (CNN) and Recurrent Neural Networks (RNN), a hybrid architecture is employed to leverage local and global contextual information. Experimental results demonstrate the superior performance of the proposed method compared to state-of-the-art techniques, highlighting its potential for improving cervical cancer diagnosis