Explainable AI Mobile App Aids Novice Sonographers in Ultrasound Diagnosis of Breast Cancer

Background: Currently, there is no artificial intelligence solution based on smartphone applications for breast cancer diagnosis. This study is the first to develop a novel model integrating original breast ultrasound images and their derived smartphone-captured photos, aiming to explore its performance in assisting radiologists with different experience levels in detecting breast cancer under a simulated real-world environment. Methods: The novel model was trained using 3,200 smartphone-captured photos derived from two datasets: 780 original breast ultrasound images from the BUSI (Breast Ultrasound Images Dataset) and 350 breast ultrasound images collected from Donghai County People's Hospital. These smartphone photos were captured by physicians following professional protocols. The model was tested on 600 externally validated smartphone photos. In a simulated real-world setting, 16 less-experienced radiologists (10 junior and 6 senior) and 5 experienced breast radiologists were invited to evaluate the model's performance using 350 prospectively collected breast ultrasound videos from 200 patients. Diagnostic performance was measured by the area under the receiver operating characteristic curve (AUC). Results: The novel model showed favorable diagnostic performance on the external validation set, with a mean AUC of 0.935 (95% CI 0.928-0.942), a sensitivity of 85.2% (95% CI 83.8-86.5%), a precision of 71.5% (95% CI 68.0-74.8%), and a kappa value of 0.721 for consistency test. In the simulated real-world test using 350 breast ultrasound videos, the model's performance was comparable to that of experienced breast radiologists (mean AUC 0.875 vs 0.890) and superior to that of junior radiologists (mean AUC 0.850 vs 0.782) and senior radiologists (mean AUC 0.842 vs 0.765). Additionally, the model improved the diagnostic performance of junior and senior radiologists: the mean AUC of junior radiologists increased from 0.782 to 0.848, and that of senior radiologists increased from 0.765 to 0.820. Conclusion: This study is the first to develop an explainable artificial intelligence-based smartphone application model, which demonstrates robust performance in breast cancer diagnosis. It can help radiologists with limited experience improve their diagnostic performance in a simulated real-world environment, providing a new technical approach for auxiliary diagnosis of breast cancer.