Fredrik Strand

Breast cancer screening reduces mortality, but disparities persist across ethnic and socioeconomic groups. This project aims to address thesse in breast cancer mammography screening outcomes across ethnic and socioeconomic groups in Sweden, particularly with the increasing use of AI systems. By analyzing approximately 380,000 retrospective screenings linked with demographics from Statistics Sweden, the project will apply AI fairness metrics to assess bias in three commercial and one in-house AI systems, conduct subgroup analyses, and use regression and multilevel modeling for statistical analysis. Stakeholder interviews and focus groups will explore perceptions of bias, and policy recommendations will be developed to enhance fairness and accessibility. The project will be conducted over three years by a multidisciplinary team from Karolinska Institutet, Lund University, and the Institute for Future Studies. Year 1 will focus on data extraction and initial assessment, Year 2 on AI evaluation and stakeholder engagement, and Year 3 on refining AI systems, pilot testing, and disseminating results. Findings will be disseminated through scientific publications, policy briefs, and training programs, guiding healthcare professionals, policymakers, and AI developers. The outcomes aim to improve screening accuracy, inform policy, and contribute to broader health equity discussions, ensuring AI supports equitable healthcare and reduces disparities.

Research problem and specific questions: Approximately 4% of middle-aged women in Sweden have breast implants, the large majority (~80%) for cosmetic reasons. At the same time there are rising concerns for adverse health effects, including reports of multiple adverse symptoms (‘breast implant illness’) as well as some psychiatric- and autoimmune conditions among women with breast implants. Yet, comprehensive research efforts are still needed to determine if these risks are preexisting or caused by the procedure itself. Therefore, the aim of the BRISK study is to leverage unique Swedish population-based data sources to determine pre- and post- implant procedure risks of: 1) mental disorders and psychotropic drug use

2) autoimmune and chronic fatigue conditions

3) sick-leave and disability pension among women with cosmetic breast implants. Data and method: We will identify 28 835 women with cosmetic breast implants through the Breast Implant Register (BRIMP

N=20 779 operated 2014-22) and ongoing mammography studies (N=8 056 with breast implants, identified 2008-21). By record linkage to the population-based Patient, Primary Care- and Prescription Medicines Registers, and Databases for Health Insurance and Labor Market, we will compare the rates of psychiatric- and autoimmune conditions as well as sick leave and disability pension among women with breast implants to that of their full sisters and an age- and region matched cohort of women (1:10 unexposed women). We will assess pre-surgery rates of these conditions and take them into account when assessing the post-surgery rates of the studied health outcomes.Societal relevance and utilization: Solid evidence on major health outcomes associated with breast implant surgery is currently lacking, leaving women considering this surgery option largely uninformed on long-term health effects. This comprehensive investigation will provide valuable information for health care policy and the growing population of women with breast implants. Plan for project realization:  We seek three years of funding for a post-doc, record linkages, database management and presentation of the results to stakeholders and the scientific community. Implementation of record linkages and preliminary analyses will be completed during the first year (2025), statistical analyses during the second year (2026), and publication of three scientific papers in leading international scientific journals by the end of the third year (2027).

Around 4% of middle-aged women have breast implants, the large majority for cosmetic reasons. There are rising concerns for adverse health effects of breast implants, yet comprehensive research efforts are needed to determine if the indicated health risks are preexisting or caused by the procedure itself. Therefore, the aim of the BRISK study is to leverage unique Swedish population-based data to determine risks of mental disorders, autoimmune disease, chronic fatigue, and other medical conditions among women with cosmetic breast implants. In this three-year project, we will identify 29 011 women with cosmetic breast implants through the Breast Implant Quality Register (operated 2014-22) and ongoing mammography studies (identified 2008-21), as well as ~5 000 women undergoing permanent removal of breast implants (1997-2022). By record linkage to the population-based Patient, Primary Care- and Prescription Registers, we will compare the rates of the studied health risks of women with breast implants to that of their full sisters and an age- and region matched cohort of women. We will assess pre-surgery rates of these conditions and take them into account when assessing post-surgery rates. Solid evidence on major health outcomes of breast implant surgery is currently lacking, leaving women considering this surgery option largely uninformed on long-term health effects. The BRISK study will provide valuable evidence for this patient group and health care policy worldwide.

Around 25% of all women who are diagnosed with breast cancer eventually die from the disease. The treatment of breast cancer patients can be improved. We have conducted retrospective studies demonstrating that AI models may reach radiologist-level performance for screening mammography. The proposed project is aiming at the diagnostic process following detection. I plan to develop AI models based on magnetic resonance imaging (MRI), which provide images richer in information compared to mammogrpahy.The project will leverage my postdoc experience with machine-learning models for breast MRI and my continued collaboration with KTH in terms of state-of-the-art AI models. In our joint AI model development for mammography, with KTH, we used convolutional neural networks. Recently, a new approach, vision transformer has been shown to be able to outperform the convolutional neural networks for image-based tasks. Therefore, my plan is to apply vision transformers to breast MRI images to address three important areas in the diagnostic process for breast cancer: image segmentation aiding radiologists to identify anatomic structures in the MRI images

radiology-pathology correlation to ascertain that biopsies correlate with image findings

therapy response prediction to inform the choice of neoadjuvant therapy. The research will tie into a recently approved EU horizon project where we will gain access to a common pool of breast cancer imaging data, and for which I am the Swedish PI.

EUropean Federation for CAncer IMages (EUCAIM) joins 76 partners to deploy a pan-European digital federated infrastructure of FAIR cancer-related de-identified images from Real-World. The infrastructure is designed to preserve the data sovereignty of providers, and provide a platform, including an Atlas of Cancer Images, for the development and benchmarking of AI tools towards Precision Medicine. EUCAIM will address the fragmentation of existing cancer image repositories by building on repositories of the AI4HI initiative, European Research Infrastructures and national/regional repositories and include clinical images, pathology, molecular and laboratory data. EUCAIM targets clinicians, researchers, and innovators, providing the means to finally build up validated clinical decision-making systems supporting diagnosis, treatment, and predictive medicine to benefit citizens. EUCAIM will define the legal grounds for the operation on a pan-European scale, adapting to the particularities of different countries on managing clinical data. EUCAIM will implement a federation of providers compliant with this legal ground, defining common data models, ontologies, quality standards, FAIR principles and de-identification procedures. EUCAIM will provide a comprehensive dashboard for data discovery, federated search, metadata harvesting, annotation and distributed processing, including federated and privacy-preserving learning. EUCAIM will build a central hub hosting the Atlas of Cancer Images, to enable development of trustworthy AI tools. EUCAIM will support new providers in building the federation and monitoring the distributed infrastructure. EUCAIM will align with the European Health Data Spaces initiative toward a sustainable flagship repository of high-quality data and tools. EUCAIM brings together clinical data providers, researchers, Research Infrastructures, and industry with mature solutions addressing the challenges of implementing such a cancer imaging infrastructure.

Breast cancer is now the most common cancer worldwide, surpassing lung cancer in 2020 for the first time. It is responsible for almost 30% of all cancers in women and current trends show its increasing incidence. Neoadjuvant chemotherapy (NAC) has shown promise in reducing mortality for advanced cases, but the therapy is associated with a high rate of over-treatment, as well as with significant side effects for the patients. For predicting NAC respondents and improving patient selection, artificial intelligence (AI) approaches based on radiomics have shown promising preclinical evidence, but existing studies have mostly focused on evaluating model accuracy, all-too-often in homogeneous populations. RadioVal is the first multi-centre, multi-continental and multi-faceted clinical validation of radiomics-driven estimation of NAC response in breast cancer. The project builds on the repositories, tools and results of five EU-funded projects from the AI for Health Imaging (AI4HI) Network, including a large multi-centre cancer imaging dataset on NAC treatment in breast cancer. To test applicability as well as transferability, the validation with take place in eight clinical centres from three high-income EU countries (Sweden, Austria, Spain), two emerging EU countries (Poland, Croatia), and three countries from South America (Argentina), North Africa (Egypt) and Eurasia (Turkey). RadioVal will develop a comprehensive and standardised methodological framework for multi-faceted radiomics evaluation based on the FUTURE-AI Guidelines, to assess Fairness, Universality, Traceability, Usability, Robustness and Explainability. Furthermore, the project will introduce new tools to enable transparent and continuous evaluation and monitoring of the radiomics tools over time. The RadioVal study will be implemented through a multi-stakeholder approach, taking into account clinical and healthcare needs, as well as socio-ethical and regulatory requirements from day one.