REPRODUCIBLE AND CLINICALLY TRANSLATABLE DEEP NEURAL NETWORKS FOR CANCER SCREENING

Area of Science:

• Oncology
• Medical Imaging
• Artificial Intelligence

Background:

• Cervical cancer is a major cause of mortality, particularly in low- and middle-income countries (LMICs).
• Scaling cervical screening in LMICs is challenging due to infrastructure and cost limitations.
• Artificial intelligence (AI) shows promise for diagnosing precancerous cervical lesions from digital images, supporting visual triage of human papillomavirus (HPV)-positive individuals.

Approach:

• A comprehensive deep-learning model selection and optimization study was conducted.
• A large, multi-institutional dataset of 9,462 women (17,013 images) was utilized.
• Model portability, repeatability, and classification performance were evaluated.

Key Points:

• The top-performing AI model achieved an area under the ROC curve (AUC) of 0.89 when combined with HPV type.
• The model demonstrated a low extreme misclassification rate of 3.4% on held-aside test sets.
• This study addresses issues of overfitting, lack of portability, and unrealistic performance estimates seen in previous AI reports.

Conclusions:

• Developed a robust, repeatable, and clinically translatable deep-learning model for cervical screening.
• AI, particularly when combined with HPV testing, can enhance the accuracy and reliability of cervical precancerous lesion detection.
• This approach offers a scalable solution to improve cervical cancer prevention in resource-limited settings.