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

Area of Science:

• Medical Image Analysis
• Computational Pathology
• Oncology

Background:

• Cervical cancer screening relies on accurate cervical cytology image classification.
• Automated systems face challenges due to subtle cellular morphology and nuclear patterns.

Purpose of the Study:

• To develop a novel deep learning architecture for enhanced cervical cytology image classification.
• To improve the accuracy and efficiency of automated cervical cancer screening.

Main Methods:

• A specialized data augmentation pipeline for cytopathology images.
• A Morphology Attention Module (MAM) for multi-scale feature extraction.
• A Spatial-Channel Mixer (SCM) for encoding nuclear spatial information.

Main Results:

• Achieved state-of-the-art accuracy: 99.06% on SIPaKMeD and 98.55% on Mendeley LBC datasets.
• Reduced error rates by up to 82.5% vs. CNNs and 61.8% vs. Vision Transformers.
• Demonstrated robust generalization across diverse cell types and imaging conditions.

Conclusions:

• The proposed deep learning model offers superior performance for cervical cytology classification.
• This technology can enhance cervical cancer screening, especially in resource-limited settings.
• Contributes to advancing automated cytology and early detection of cervical abnormalities.