Background: Diagnostic mammography with reasonable sensitivity and specificity are currently the two primary imaging modalities for early detection of breast cancer. To assist radiologists, breast physicians and registrars in improving diagnosis on mammograms, the BREAST (Breastscreen REader Assessment STrategy) program has been implemented nationally to assist readers in reviewing mammograms with instant results and meaningful feedback after completing online test sets [1,2]. Our study aims to use artificial intelligence (AI) to enable readers to interact with individualizing training sets to enhance readers’ performances in breast cancer detection.
Methods: We recruited three participants (senior radiologist, junior radiologist and breast physician) to undertake AI training sessions of reading digital mammograms on the BREAST platform. The participants completed two BREAST test sets (120 cases) in the last three years. The raw data of readers were analysed to generate the diagnostic efficacy of each participant in specificity, case sensitivity and lesion sensitivity. Based on analyzed strengths and errors, case-based exercises with different cancer features were assigned to each participant in form of training sets. The difficulty coefficient for every exercised case was calculated based on the performances with more than 194,000 data points of 586 BreastScreen readers in BREAST databank using Bayesian method. As the training goes forward, the cases with higher level of difficulty were selected to improve the reader diagnosis in more complicated cases. The performances of participants after training sessions were compared with the previous results to evaluate the effectiveness of the training sessions.
Results: Two participants completed one training session, while another participant completed two sessions. On average, the participants exhibited a noteworthy 13% increase in case sensitivity, rising from 79% to 89%. Their improved capability to detect the accurate location of abnormalities was also evident, with a 10% increase in lesion sensitivity from 74% to 80%. Although the readers' sensitivity displayed significant improvement, their specificity remained relatively consistent, averaging between 85% and 83% before and after the training sessions for all three participants.
Conclusion: The findings suggest that the readers have enhanced their ability to correctly identify cancer cases in screening mammograms while maintaining reasonable specificity in identifying non-cancerous cases during the training sessions. To conduct more robust data analysis pertaining to the training sessions, we are in the process of recruiting additional participants and organizing more training sessions. By continuously refining our approach, we strive to make significant contributions to breast screening programs and ultimately enhance patient outcomes.