Multicontrast Pocket Colposcopy Cervical Cancer Diagnostic Algorithm for Referral Populations

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Cervical cancer causes significant mortality, particularly in low-resource settings.
• Shortage of trained professionals and diagnostic variability hinder early detection.
• Automated classification algorithms can address these challenges.

Purpose of the Study:

• To develop and evaluate deep learning models for classifying high-grade cervical precancer and cancer.
• To enhance classification performance using class-balanced loss and green-light colposcopy images.
• To improve early detection of cervical lesions in resource-limited areas.

Main Methods:

• Utilized a dataset of cervical images from 880 patient visits.
• Optimized deep learning network architecture and incorporated a weighted loss function.
• Explored the integration of green-light colposcopy image pairs to improve model sensitivity.

Main Results:

• Achieved an area under the receiver-operator characteristic curve of 0.87.
• Reported a sensitivity of 75% and specificity of 88%.
• Class-balanced loss and green-light imaging improved sensitivity by 2.5 times.

Conclusions:

• The developed methodology enhances classification performance for cervical precancer and cancer detection.
• This approach can be integrated with existing screening methods like Pap smears or HPV testing.
• Broadens access to early detection, potentially reducing cervical cancer mortality.