N1-0197 Improving breast cancer screening through dynamic big data analytics of Quantitative Imaging Biomarkers (QIBs)
Research project is (co) funded by the Slovenian Research Agency.
UL Member: Faculty of Mathematics and Physics
Project: Improving breast cancer screening through dynamic big data analytics of Quantitative Imaging Biomarkers (QIBs)
Period: 1. 1. 2021 - 31. 12. 2024
Range per year: 1,2 FTE, category: B
Head: Robert Jeraj
Research activity: Natural sciences and mathematics
Citations for bibliographic records
Breast cancer is an important health problem, successfully managed by screening programs. Slovenian and Flemish programmes meet the strict criteria dictated by EU documents. In the recent years, two critical factors have been realized:
- there are five times as many interval cancers being discovered among population with dense breast compared to other breast density participants
- one fifth of cancers discovered by mammography is present in the mammogram recorded at the preceding screening visit
Breast density and tissue masking are different breast characteristics, both important and both contributing to the breast cancer risk. These risks are seldom evaluated in screening program due to limited knowledge of reliability of image derived measures.
To provide reliable measures, imaging needs to be elevated from quantitative reads to "Quantitative Imaging Biomarkers (QIBs)". By definition QIBs need to be objectively derived from images and measured on a ratio or interval scale of normal biological processes, pathogenic processes, or a response to a therapeutic intervention. QIBs can only be reliably extracted through advanced image analytics.
Image analytics uses methods based on (1) individual voxel intensity, (2) voxel location relative to other landmarks, or (3) local patterns around each voxel or any combination thereof. Such combinations are referred to as image filters, which can be either hand-crafted or machine-learned from data. Hand-crafted filters include radiomics textures and leverage prior knowledge of expected patterns related to condition with condition intensity - tissue granularity, common patterns - and are easier to interpret, implement and adjust, but are often not optimal and/or flexible for different tasks. Machine learned features are statistically more significant and can be used for different tasks, but are difficult to interpret and require significant computer power to learn.
Our goal is to improve breast cancer screening through dynamic big data analytics of Quantitative Imaging Biomarkers (QIBs) that fully assess all available breast characteristics spatially and temporarily. In other words, we want to integrate information on:
- dynamic changes from one screening time point to another (longitudinal assessment) and
- big data analytics extracting features predictive of cancer (spatial assessment)