Cervical cancer is the second most common cancer among women in Cambodia and the second leading cause of cancer death in women aged 15 to 44, despite a low HPV prevalence of around 5%. The gap is largely one of access: diagnostic tools and vaccines are expensive and often out of reach in resource-limited settings.
Neural-network screening for digital colposcopy is a promising answer, but it runs into a data problem. Training on images from a country with low HPV prevalence produces a badly imbalanced dataset in which the high-grade lesion cases, the ones that matter most because they carry the highest risk of progressing to cancer, are scarce. The usual remedies all have costs: oversampling the rare class invites overfitting, and undersampling the common one discards information that small medical datasets can ill afford to lose.
What we did
Rather than resampling what already exists, we generate what is missing. A diffusion model was trained to synthesise high-grade lesion features within otherwise normal cervical images, producing realistic positive cases instead of duplicating the few that were available.
What we found
The method was evaluated on datasets from Berlin and from Cambodia, the latter being the one with the pronounced shortage of high-grade lesions. The approach balanced the dataset and improved diagnostic performance by 5%.