Deep Learning-Based Recognition of Cervical Squamous Interepithelial Lesions

Area of Science:

• Digital pathology
• Artificial intelligence in oncology
• Cervical cancer precursor detection

Background:

• Cervical squamous intraepithelial lesions (SILs) are critical precursors to cervical cancer, necessitating accurate diagnosis for timely intervention.
• Current diagnostic methods for SILs in histology are often laborious and suffer from low inter-observer consistency due to image similarities.
• Existing artificial intelligence (AI) models for cervical histology lack robust feature extraction and do not adequately incorporate p16 immunohistochemistry (IHC) data.

Purpose of the Study:

• To develop and validate an AI-driven approach for the accurate classification of cervical squamous intraepithelial lesions (SILs) using histology and p16 IHC data.
• To improve the diagnostic accuracy and consistency of identifying high-grade squamous intraepithelial lesions (HSIL).
• To provide a tool that assists pathologists in diagnosing cervical lesions and informs patient treatment strategies.

Main Methods:

• A squamous epithelium segmentation algorithm was developed and applied to label relevant regions.
• Whole Image Net (WI-Net) was utilized to extract p16-positive areas from IHC slides.
• p16-positive masks were generated by mapping IHC data back to H&E slides, serving as input for Swin-B and ResNet-50 deep learning models for HSIL classification.

Main Results:

• The Swin-B model achieved an accuracy of 0.914 for HSIL detection.
• The ResNet-50 model demonstrated an AUC of 0.935, with an accuracy of 0.845, sensitivity of 0.922, and specificity of 0.829 at the patch level for HSIL.
• The developed AI models showed high performance in classifying HSIL from cervical histology images.

Conclusions:

• The proposed AI model, particularly the Swin-B and ResNet-50 architectures, can accurately identify high-grade squamous intraepithelial lesions (HSIL).
• This AI tool has the potential to assist pathologists in clinical diagnosis, improving efficiency and consistency.
• The integration of p16 IHC data within the AI framework enhances the identification of cervical precursor lesions, potentially guiding patient management.