AI integration improves breast cancer screening in a real-world, retrospective cohort study

Artificial intelligence (AI) holds promise for improving breast cancer screening, but many challenges remain in implementing AI tools in clinical screening services. AI readers compare favourably against individual human radiologists in detecting breast cancer in population screening programs. However, single AI or human readers cannot perform at the level of multi-reader systems such as those used in Australia, Sweden, the UK, and other countries. The implementation of AI readers in mammographic screening programs therefore demands integration of AI readers in multi-reader systems featuring collaboration between humans and AI. Successful integration of AI readers demands a better understanding of possible models of human-AI collaboration and exploration of the range of possible outcomes engendered by the effects on human readers of interacting with AI readers. Here, we used a large, high-quality retrospective mammography dataset from Victoria, Australia to conduct detailed simulations of five plausible AI-integrated screening pathways. We compared the performance of these AI-integrated pathways against the baseline standard-of-care "two reader plus third arbitration" system used in Australia. We examined the influence of positive, neutral, and negative human-AI interaction effects of varying strength to explore possibilities for upside, automation bias, and downside risk of human-AI collaboration. Replacing the second reader or allowing the AI reader to make high confidence decisions can improve upon the standard of care screening outcomes by 1.9-2.5% in sensitivity and up to 0.6% in specificity (with 4.6-10.9% reduction in the number of assessments and 48-80.7% reduction in the number of reads). Automation bias degrades performance in multi-reader settings but improves it for single-readers. Using an AI reader to triage between single and multi-reader pathways can improve performance given positive human-AI interaction. This study provides insight into feasible approaches for implementing human-AI collaboration in population mammographic screening, incorporating human-AI interaction effects. Our study provides evidence to support the urgent assessment of AI-integrated screening pathways with prospective studies to validate real-world performance and open routes to clinical adoption.