A deep learning approach for Cervical Automated Risk Assessment (CARE) using images from a low-cost, portable Pocket colposcope

For cervical cancer screening in low-HDI countries, the WHO recommends that a diagnosis is made immediately upon cervical visualization. To address variability in provider visual interpretations, we use CNNs to classify images from a low-cost, FDA-certified, portable Pocket colposcope images positive for high-grade precancer from a triaged population. We show that the combination of white-light acetic acid and green-light image stacks improves the AUC to 0.9. Pocket CARE can be used at the community level without the need for specialized physicians or inaccessible equipment, broadening access to early detection and treatment of pre-cursor lesions before they advance to cancer.