Enhancing Cervical Pre-Cancerous Classification Using Advanced Vision Transformer

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Cervical cancer incidence and mortality are rising globally, particularly in developing nations.
• Limited access to screening, specialists, and awareness exacerbates the problem.
• Current screening methods include visual inspection with acetic acid (VIA), histopathology, Papanicolaou (Pap) tests, and human papillomavirus (HPV) tests.

Purpose of the Study:

• To develop an automated system for cervical pre-cancer identification using artificial intelligence.
• To enhance the accuracy and accessibility of cervical cancer screening.
• To create a robust decision support tool for clinicians, especially in low-resource settings.

Main Methods:

• Utilized a vision transformer (ViT) model enhanced with shifted patch tokenization (SPT).
• Trained and tested the model on 8215 colposcopy images from the mobile-ODT dataset.
• Evaluated the model's generalization capability on 30% of the dataset.

Main Results:

• Achieved 91% accuracy in identifying three distinct cervical pre-cancer types.
• Demonstrated superior performance compared to state-of-the-art methods.
• The model effectively learned distinguishing features between cervical pre-cancerous types.

Conclusions:

• The developed ViT-SPT system offers a highly accurate and automated approach for cervical pre-cancer detection.
• This technology can serve as a valuable decision support tool, improving screening in underserved regions.
• The system holds significant potential for early detection and management of cervical cancer globally.