Impact of Image and Feature Harmonization on the Computerized Analysis of Mammograms

One of the main challenges faced by medical image analysis algorithms is their reduced generalization capability to unseen data. In the automated analysis of mammograms for breast cancer research, this problem can been tackled by the incorporation of image pre-processing or feature post-processing harmonization techniques. This research aims to assess the impact of these techniques on the performance of machine learning algorithms for breast cancer risk assessment. We conduct a retrospective case-control study on 147 mammograms acquired with mammographic systems from two different vendors and consider the impact of harmonization techniques in three different experimental settings: 1) Intra-vendor, when all images are from the same vendor, 2) Inter-vendor, when training and testing images are captured with systems from different vendors, and 3) Mixed-vendor when training and testing include images from both vendors. Our results show that the joint use of image and feature harmonization is beneficial in the inter-vendor setting but detrimental in the intra-vendor setting. The mixed-vendor setting does not show significant improvement with harmonization.