Breast cancer is the single most diagnosed type of cancer among women worldwide. Fortunately,mammography screening successfully decreases breast cancer mortality, yet entails a number of prac-tical challenges related to cumulative radiation dosage, overdiagnosis, breast radiologists’ workload,lack of breast radiologists, anxiety for screened/biopsied women, and differences in screening qualityfrom country to country.Personalized breast cancer screening, based on individual risk, is proposed to alleviate the men-tioned practical challenges associated with large scale screening. Traditional risk models based onclinical risk factors are in use today, but are practically limited by continuous extensive data collectionprocesses. Mammography-based risk models, that solely consider the imaging data, are more suitedfor clinical implementation in screening, as the mammograms are readily available. Mammography-based risk models have been studied for decades and have gradually improved, however, deep learn-ing now offers methods of reliably estimating risk with high accuracy, robustness, and objectivity.Mammography-based deep learning models for detection are commercially available and can de-tect cancer with the same accuracy as radiologists. In the first study, we employed an (artificialintelligence) AI system for lesion detection and retrospectively simulated a screening protocol inwhich women were categorized based on the likelihood of detecting a breast cancer. Women withlikely normal mammograms were excluded from radiologist reading as healthy. Women with veryhigh likelihood of breast cancer were recalled immediately for diagnostic tests. Women with mod-erate likelihood was double-read as in standard screening. The results indicated that, with AI-basedscreening, the radiologists’ workload could be considerably decreased while the high screening qual-ity was preserved. This simulation study contributed to a full clinical implementation of the AI systemin the Capital Region of Denmark for which prospective preliminary results indicate a safe rollout ofAI in screening.Recent studies have shown promise in estimating risk of a future breast cancer. However, currentrisk models does not ensure reliable risk assessment across mammographic devices from differentvendors. Additionally, current risk models are trained in a conflated manner and learn features in-dicative of a visible breast cancers, subtle signs of developing breast cancers, and textural featuresindicative of a susceptibility to future breast cancer. Training for all three tasks simultaneously mightlead to subpar long-term risk assessment. In the second study, we developed a texture model that wastrained optimally for long-term risk, relying on features of healthy breast tissue indicative of futurebreast cancer. We demonstrated that the texture model could robustly estimate short- and long-termrisk while generalizing across mammographic devices from different vendors.We additionally developed a dense tissue segmentation tool to estimate planimetric percentagemammographic density (PMD), which is a known and established breast cancer risk factor.In the third study, we combined the AI system for short-term risk and the texture model forlong-term risk in a combination model with age and PMD to create a rich mammography-based risk profile. The results indicated that training each system individually and combining them subsequentlysignificantly improved overall risk assessment.In the fourth study, we developed a generic machine learning framework for identifying patientgroups using national health registry data. This framework was applied to identification of breastcancer relapse patients, but could be directly translated to breast cancer risk assessment and used foran even more expressive breast cancer risk profile.These developed models might, with further rigorous validation, support clinicians in creatingpersonalized screening protocols that could benefit patients and radiologists.