Artificial intelligence and visual inspection in cervical cancer screening

Area of Science:

• Gynecology and Obstetrics
• Medical Imaging
• Artificial Intelligence in Healthcare

Background:

• Visual inspection with acetic acid (VIA) for cervical cancer screening is subjective and resource-intensive.
• Limitations of VIA include variability in interpretation and a shortage of skilled personnel.
• Artificial intelligence (AI) offers a potential solution to enhance the objectivity and accessibility of VIA.

Purpose of the Study:

• To evaluate the diagnostic performance of healthcare workers, experts, and an AI algorithm in visual inspection with acetic acid (VIA).
• To compare the accuracy of AI and human interpretation against expert consensus for cervical cancer screening.
• To assess the feasibility of using expert consensus as a reference standard for training and AI development.

Main Methods:

• A diagnostic study involving 22 healthcare workers and 9 gynecologist experts interpreting 83 cervical images.
• An AI algorithm was also used to assess the same image set.
• Diagnostic performance metrics included sensitivity, specificity, area under the curve (AUC), and inter-observer agreement (Fleiss kappa).

Main Results:

• Experts achieved the highest diagnostic performance (Sensitivity: 81.6%, Specificity: 93.5%, AUC: 0.93).
• The AI algorithm demonstrated strong performance (Sensitivity: 80.0%, Specificity: 83.3%, AUC: 0.84), comparable to healthcare workers (Sensitivity: 80.4%, Specificity: 80.5%, AUC: 0.80).
• Inter-observer agreement was highest for experts (kappa=0.68) and the AI algorithm (kappa=0.63) compared to healthcare workers (kappa=0.45).

Conclusions:

• Expert consensus serves as a viable reference standard, potentially replacing histopathology for training purposes.
• AI algorithms show significant potential in improving the accuracy and consistency of cervical cancer screening via VIA.
• The study highlights the complementary roles of AI and expert judgment in enhancing diagnostic accuracy and addressing resource limitations in cervical cancer screening.