The investigation of AI models on cancer sizes of screening mammograms

Background and purpose: Breast cancer is the most common cancer among Australian women, with over 20,000 Australians diagnosed with breast cancer each year (representing 28% of all reported cancers in females)1. Artificial Intelligence (AI) has been widely used in medical imaging analysis due to providing promising results in the last decade2. This study aims to evaluate the diagnosis accuracy of AI systems on mammograms with different groups of cancer sizes. Methods: The cancer sizes were reported in histopathological report with 856 biopsy-proven cancer on screening mammograms. The Globally-aware Multiple Instance Classifier (GMIC) and Global-Local Activation Maps (GLAM) models2 were used to detect breast cancer on screening mammograms. Transfer learning of GMIC and GLAM models was conducted using 1712 mammograms (856 cancer and 856 normal) from an Australian database. The cancer sizes were classified as three groups: T1 (<=2cm), T2(>2cm & <5cm), and T3(>=5cm) according to the American Cancer Society. The diagnosis performance of the AI models was evaluated in terms of sensitivity on mammograms with different groups. The ANOVA test was conducted for statistical analysis of the AI models across different size groups. Results: GMIC outperformed GLAM in all three groups of cancer sizes in terms of sensitivity. The sensitivity of GMIC models on T1, T2 and T3 was 0.882, 0.928, and 0.947, respectively. GMIC (P-value=0.022) demonstrated significant difference in sensitivity across three groups of cancer sizes. The sensitivity of GLAM models on T1, T2 and T3 was 0.851, 0.894, and 0.923, respectively. Similarly, GLAM (P-value=0.016) demonstrated significant difference in sensitivity across three groups. Conclusions: The diagnosis accuracy of AI systems is affected by the cancer sizes in mammograms, with larger size of cancer yielding more accurate cancer detection. The findings of this study indicate that the AI models can provide accurate detection of cancers on mammograms, particularly when the cancer sizes are sufficiently large. References 1. Australian Institute of Health and Welfare, Breast cancer in Australia: an overview, AIHW, Australian Government, (2012). 2. Stadnick B, Witowski J, Rajiv V, Chłędowski J, Shamout FE, Cho K, Geras KJ. Meta-repository of screening mammography classifiers. arXiv preprint arXiv:2108.04800, (2021).