Deep learning models for cervical cancer subtyping using whole slide images.

Area of Science:

• Computational pathology
• Artificial intelligence in oncology
• Digital pathology

Background:

• Cervical cancer subtyping is crucial for effective treatment strategies.
• Accurate subtyping can be challenging with traditional methods.
• Whole-slide images (WSI) offer rich data for computational analysis.

Purpose of the Study:

• To develop and evaluate an artificial intelligence (AI)-based model for cervical cancer subtyping.
• To improve diagnostic accuracy using both patch-level and WSI-level analyses.
• To assess the model's generalization performance on independent datasets.

Main Methods:

• Utilized 438 whole-slide images from public and private datasets.
• Employed a two-stage approach: patch-level prediction and WSI-level prediction.
• Applied convolutional neural networks (CNNs) for patch-level analysis and machine learning algorithms (e.g., SVM with TF-IDF) for WSI-level analysis.

Main Results:

• Inception-v3 achieved high AUROC (0.960) at the patch level.
• Support Vector Machine (SVM) with TF-IDF features demonstrated superior WSI-level performance (AUROC up to 0.964).
• The model exhibited strong discrimination and calibration, validated by decision curve analysis.

Conclusions:

• AI models show significant potential for accurate cervical cancer subtyping.
• The developed model demonstrates robust generalization across different datasets and clinical settings.
• This approach holds promise for enhancing cervical cancer diagnosis and patient management.