Purpose:
The goal of this work is to develop a method to assess the risk of breast cancer masking, based on image characteristics beyond breast density.
Method:
From the Dutch breast cancer screening program we collected 285 screen detected cancers, and 109 cancers that were screen negative and subsequently appeared as interval cancers. To obtain mammograms without cancerous tissue, we took the contralateral mammograms. We developed a novel machine learning based method called convolutional sparse autoencoder to characterize mammographic texture. The method was trained and tested on raw mammograms to determine cancer detection status in a five-fold cross validation. To assess the interaction of the texture scores with breast density, Volpara Density Grade was determined for each image.
Results:
We grouped women into low (VDG 1/2) versus high (VDG 3/4) dense, and low (Quartile 1/2) versus high (Q 3/4) texture risk score. We computed odds ratios for breast cancer masking risk (i.e. interval versus screen detected cancer) for each of the subgroups. The odds ratio was 1.63 (1.04-2.53 95%CI) in the high dense group (as compared to the low dense group), whereas for the high texture score group (as compared to the low texture score group) this odds ratio was
2.19 (1.37-3.49). Women who were classified as low dense but had a high texture score had a higher masking risk (OR 1.66 (0.53-5.20)) than women with dense breasts but a low texture score.
Conclusion:
Mammographic texture is associated with breast cancer masking risk. As such, automatic texture analysis offers opportunities to enhance personalized breast cancer screening. We were able to identify a subgroup of women who are at an increased risk of having a cancer that is not detected due to textural masking, even though their breasts are non-dense.