DeepCIN: Attention-Based Cervical histology Image Classification with Sequential Feature Modeling for...

Area of Science:

• Computational pathology
• Digital histopathology
• Machine learning for medical diagnosis

Background:

• Cervical cancer poses a significant global health threat to women.
• Histopathological examination of cervical biopsy slides for cervical intraepithelial neoplasia (CIN) grading is prone to interobserver variability.
• Automated analysis of digitized histopathology slides offers potential for enhanced accuracy in classifying CIN grades (Normal, CIN1, CIN2, CIN3).

Purpose of the Study:

• To develop and evaluate DeepCIN, a hierarchical network pipeline for automated CIN grading.
• To model the progression of cervical disease within epithelial tissue using spatial information.
• To improve the accuracy and consistency of CIN classification compared to traditional methods.

Main Methods:

• A two-stage deep learning pipeline, DeepCIN, was developed for analyzing high-resolution epithelium images.
• The pipeline employs a vertical segment-level sequence generator using weak supervision to capture bottom-to-top feature relationships.
• An attention-based fusion network integrates local segment information for final image-level CIN grade prediction.

Main Results:

• The DeepCIN model successfully produced CIN classification results.
• The system identified specific vertical segments contributing to the CIN grade predictions.
• The model demonstrated its capability to analyze spatial disease progression within epithelial tissue.

Conclusions:

• DeepCIN achieves classification accuracy comparable to that of experienced pathologists.
• The automated approach holds promise for more objective and reliable CIN grading.
• This technology could significantly enhance cervical cancer screening and diagnosis.