Background:

Mammography is the most common imaging modality for the detection of breast cancer. Artificial intelligence algorithms for mammography analysis have shown promising performance for breast cancer risk assessment and lesion detection and classification; however, these models often fail the test of external validation. The evidence points to variations in image acquisition-known as the batch effect-as a main contributing factor to the lack of the models generalization and robustness. However, studies on the effects of acquisition in the mammogram have been limited due to lack of appropriate datasets. This prospective, exploratory, non-randomized study aims to study how biological and non-biological sources of heterogeneity affect the mammogram and, in turn, the computerized models for mammography analysis.

Methodology:

This study will collect breast samples from 200 participants that will undergo breast resection as per clinical indications. Each sample will undergo the mammography imaging procedure several times to obtain mammograms with different combinations of imaging parameters. The resulting dataset will be used for the statistical analysis of the impact of imaging parameters in mammographic texture features and the computerized analysis of mammograms. Furthermore, biological information will be collected from the resected breast samples to study their relation to mammographic texture features.

Discussion:

This study will add to the understanding of the effect of different sources of heterogeneity on mammography, ultimately aiding in the future development of robust computerized analysis models.