DEEP-BREAST

Breast cancer is the most frequently diagnosed cancer in women worldwide. A malignancy in the breast is often associated with local stiffness changes that occasionally can be detected by manual palpation. However, what the doctor feels during palpation does not always correspond to what the radiologist sees in mammography images of the breast. This hampers reliable detection, localization and characterization of suspected lesions seen in mammography and often requires extensive and expensive additional diagnostic procedures like Ultrasound and MRI to come to a final diagnosis. 

In DEEP-BREAST we propose to measure mechanical parameters of the breast during a mammographic examination. This allows establishing a direct link between what is seen in a 3D mammography (tomosynthesis) image of the breast and corresponding quantitative measurement of local stiffness at the same location. These mechanical parameters are expected to yield information that has the potential to increase detectability and characterisation of breast lesions. 

To estimate mechanical parameters of the breast we propose two approaches: 1) analysing data acquired with the Sigma Paddle, a special compression paddle developed by Sigma screening that enables recording mechanical and geometrical parameters (breast thickness, force, pressure, contact area) during breast compression. 2) measuring local strain in the breast obtained from two tomosynthesis images at different compression levels. These parameters will be used for the development and training of a machine learning application (deep learning) to establish the relationship between mechanical information and breast pathology.  

We envision that, after successful completion of the project, knowledge will be available to develop a deep learning application that allows improvement of breast cancer detection and characterization based on mechanical parameters. This application will direct the application of the Sigma Paddle and associated analysis software towards a product that enables improved and cost effective breast diagnostics in a single modality. 

The deliverables of the project were development and evaluation of methods for breast stiffness estimation from the sigma paddle recordings as well as from strain patterns obtained from two tomosynthesis  images.  

We performed in-depth manual exploration on finding possible combinations of measured and derived metrics from the paddle data as biomarker for pathology, as well as machine learning approaches for predicting pathology. The evaluations show that paddle data has some predictive value for pathology. Fundamental knowledge is gained concerning strain estimation in breasts from 3D imaging. We discovered that the characteristics of the data sets of the tomosynthesis system was not yet suitable to reliably estimate strain patterns and correlate them to pathology.