Multi-scale feature integration with enhanced cytomorph for high-accuracy cervical cytology classification

Accurate classification of cervical cytology images plays a crucial role in early detection and prevention of cervical cancer, which remains a significant global health challenge. Despite advancements in deep learning for medical image analysis, the unique characteristics of cervical cells, including subtle morphological differences and complex nuclear patterns, pose considerable challenges for automated classification systems. In this paper, we present a novel deep learning architecture specifically designed for cervical cytology image classification. Our approach integrates three key components: (1) a specialized data augmentation pipeline tailored for cytopathology images, (2) a Morphology Attention Module (MAM) that captures multi-scale cellular features with adaptive feature fusion, and (3) a Spatial-Channel Mixer (SCM) that efficiently encodes nuclear neighborhood spatial information. Extensive experiments on both the SIPaKMeD and Mendeley LBC datasets demonstrate the superior performance of our model, achieving state-of-the-art accuracy of 99.06% on the 5-class SIPaKMeD dataset and 98.55% on the Mendeley LBC dataset. Importantly, our approach reduces error rates by up to 82.5% compared to conventional CNN architectures and 61.8% compared to recent Vision Transformer approaches. The proposed architecture demonstrates robust generalization across different cell types and imaging conditions, making it a promising tool for enhancing cervical cancer screening programs. Our work contributes to the advancement of automated cytology analysis and has the potential to improve early detection of cervical abnormalities, particularly in resource-limited settings where expert cytopathologists may be scarce.