In an NCI-supported study, researchers applied artificial intelligence (AI) to cervical cancer screening to determine how well an AI model could perform across six different countries while comparing images from a camera and smartphone. This study emphasizes the importance of considering data variety when designing AI models. The study shows how you can generate and externally validate a cervical cancer diagnostic classifier that is reliable, consistent, accurate, and clinically translatable. The results, when applied, can also help better triage women into appropriate cancer risk categories. Read on for a summary of the results as well as a link to the full publication.

The team used their image recognition model, which classifies images into normal, intermediate, and precancer categories. The researchers found that the model faced more challenges in predicting cancer status when presented with images from different devices versus different geographies from the training data. Researchers also observed improved performance when they retrained the model with captured images from the smartphone.

From their research, you can get valuable insights into emerging trends and methodologies in data analysis. Retraining the model to improve its adaptability to different data sources highlights key considerations for:

• developing robust AI models.
• assessing the relative impact of various axes of diverse data on model performance.