| Abstrakt: |
Introduction: Breast cancer (BC) is a global concern affecting 1 in 8 women with 2.3 million new cases and 685,000 deaths in 2020. Risk prediction and early detection are crucial for BC management. Risk prediction models mainly consider factors like lifestyle, genetics, age, body mass index, menopausal situation, family history, microcalcifications, breast masses, and mammographic density. Radiomic features, especially mammographic density, hold promise among these features. Artificial intelligence (AI) models, particularly convolutional neural networks (CNNs) using radiomics, enhances BC risk assessment, outperforming traditional methods. This study aims to review the role of AI models in improving BC risk prediction. Methods: PubMed, Science Direct, Web of Science, and Google Scholar databases were explored up to May 2023, using different combinations of the keywords: "Breast Cancer", "Deep - Learning", "Mammographic Density", "Convolutional Neural Networks", "Artificial Intelligence", "Predictive Models" and "Risk Assessment". Eight more recent and relevant records were included in the study. Results: The findings from the reviewed studies on BC risk prediction have demonstrated diverse accomplishments. BC risk prediction models based on CNN architecture for extracting mammographic radiomics achieved a pooled accuracy of about 70% through pixel - wise matching and consideration of breast density. Notably, an Inception ResNet - v2 deep learning (DL) model surpassed density - based models in predicting 5 - year BC risk, particularly in cases involving aggressive tumors. Additionally, unsupervised DL model of convolutional sparse autoencoder (CSAE) network, exhibiting promising results in BC risk prediction (AUC = 0.59). However, some studies have shown the accuracy up to about 86% for hybrid models that benefit from logistic regression, CNN-based models derived radiomics and other BC risk factors. Conclusions: Deep learning, despite being a black box, provides superior accuracy, aiding radiologists in screenings. Locally validated networks aim to create a specific screening tool for targeted risk assessment, reducing unnecessary mammography and identifying high - risk patients. The reviewed records collectively highlight the potential of AI models in improving BC risk prediction. While results vary among different studies, the integration of clinical factors with radiomics extracted from advanced imaging modalities using DL models, holds promise for personalized BC risk assessment and improved prediction accuracy. [ABSTRACT FROM AUTHOR] |
