**3.6 Hematology/oncology**

The utilization of artificial intelligence (AI) in cancer screening is becoming increasingly evident in recent studies across multiple types of cancer. This includes lung, breast, colorectal, and cervical [47–50]. Given the overwhelming research across multiple disciplines, the focus of this review will be on the evidence-based application of AI in breast cancer screening. This research can be categorized into two applications: risk assessment and image analysis.

The United States Preventive Services Task Force (USPSTF) guidelines of primary screening for breast cancer with conventional mammography has resulted in a reduction of breast cancer mortality across both randomized trials and screening cohort studies [51]. Outlined in the USPSTF recommendations is screening every 2 years for women aged 50–74 years old, as opposed to individualized decision to start screening between the ages of 40–49 years old [51]. In the latter age group, high-risk individuals who would benefit from starting screening at an earlier age can include those with known underlying genetic mutation (such as BRCA1 or BRCA2 gene mutation) or a

history of chest radiation at an early age [51]. There are several risk prediction models for breast cancer. One example is the Breast Cancer Risk Assessment Tool (BCRAT), which can be used to estimate a patient's 5-year and lifetime risk of developing invasive breast cancer. This considers a patient's age, age of menarche, age of first childbirth, number of first-degree relatives with breast cancer, number of previous biopsies, and presence of atypical hyperplasia in a biopsy. Of note, this tool may not be appropriate for assessing risk in patients with a history for certain medical conditions, such as personal history of certain breast cancer types [52]. Considering multiple qualitative and quantitative risk factors can better stratify risk-based screening and maximize the benefit while minimizing the harms of screening [53]**.** But how can AI advance current tools of risk assessment?

Breast density has been shown to be an independent risk factor for the development of breast cancer [54]. As a result, this has led to updates in prediction models to include this quantitative risk factor, such as the Tyrer Cuzick model, the Breast Cancer Surveillance Consortium Model (BCSCM), and the Breast and Ovarian Analysis of Disease Incidence and Carrier Estimation Algorithm (COADICEA) [52]. In a recent study, the authors created three models to estimate five-year breast cancer risk. One model only considered risk factors. The second model utilized deep/machine learning on mammographic images. The third model was a hybrid of the two. These models were then compared to the Tyrer–Cuzick model, a well-known clinical standard that recently incorporated mammographic breast density into its calculation. They found that their hybrid model had the highest accuracy, followed by the deep/machine learning model, while the Tyrer–Cuzick model had the lowest. These results indicate that a model that considers both traditional risk factors and mammographic data can improve current practices of assessing risk. Future research can aim to identify the imaging features and patterns that are most useful to stratifying risk [52].

Breast density is typically assessed through interpretation of the standard twoview mammogram by a radiologist. A visual estimation of the proportion of glandular and fatty tissue within the breasts is scored and applied to a scale, such as the Breast Imaging Reporting and Data System (BI-RADS). The four BI-RADS categories of breast composition according to breast density are: type 1 fatty breast, type 2 fibroglandular, type 3 heterogeneously dense, and type 4 dense and homogeneous. This subjective quantification of breast density requires certain training and experience to allow for accurate and reproducible scoring. Even so, there is a certain amount of user variation among radiologists that contributes to error [55].

There are 3 potential approaches to applying AI to mammogram image analysis: as a standalone system, for triage, and for reader aid [56]. In a simulation performed by McKinney et al., the findings demonstrated the ability of an AI system to outperform a group of radiologists in accurately interpreting mammograms [57]. Using deep learningbased AI, Balta et al. found that the breast cancer screening workflow, which typically requires double-reading, could be replaced by a single-reading. This was achieved by AI-driven identification of normal-appearing screening mammograms, which were then verified by a single human reader [56]. Similarly, in a retrospective study by Dembrower et al., AI was used to triage mammograms into those requiring no further radiologist assessment and those requiring further radiologist assessment. This system demonstrated potential for detecting a significant number of cases where breast cancer was not identified by human readers, but then diagnosed later [58]. Rodriguez-Ruiz et al. showed that radiologists interpreting mammograms with the support of an AI computer system performed better at diagnosing breast cancer than without [59].
