The most commonly detected cancer type worldwide in 2020 was Breast Cancer (BC). Early diagnosis of this disease, aimed through generalized screening programs with mammography, is imperative to improve BC prognosis. Despite its positive impacts, these programs present some pitfalls. The two-dimensional nature of screening mammography often results in tissue overlap, which can obscure the presence of tumors. This phenomenon contributes to false negative results, potentially delaying cancer diagnosis and compromising disease prognosis. This work proposes an Artificial Intelligence (AI) model capable of analyzing current healthy mammograms and predicting their masking potential. This refers to the likelihood of a potential future cancer, if present, being obscured in subsequent screening mammograms. Given that, 3,000 synthetic mammograms, evenly divided into three masking potential classes Low, Medium, and High were used to train a Convolutional Neural Network (CNN). The performance of the CNN was evaluated using a test set comprising 1,000 mammograms from each masking potential class. Besides that, an independent test set comprised of real instead of synthetic mammograms (N = 201) was also used to assess performance. The F1-score, Specificity, and Accuracy values were very high on the synthetic test set, measuring at 0.976, 0.988, and 0.976, respectively, underscor ing the excellent predictive capability of the CNN. Moreover, the results on the independent test set also show a high classification capacity on the Low and High masking classes in terms of Precision and Specificity. A model like the one proposed here can have significant impacts in the future, allowing personal ized screening based on masking risk, potentially reducing the number of false negative results and ultimately improving the outcomes of this disease.
Read full abstract