Despite being one of the leading cancers in women, breast cancer detection rates in a repeat screened population are quite low (i.e., 3 to 5 cancers detected per 1000 examinations). Screening for the early detection of breast cancer has been controversial from the start, but recent events highlight the need to develop and optimize individualized screening regimens by identifying women who are at higher than average risk of developing breast cancer in the near future, namely within five years. Establishing optimal individualized screening regimens that facilitate women to be screened at different intervals and/or with different imaging methods based on their assigned risk group will not only increase sensitivity, resulting in the detection of earlier cancers, but also reduce overall cost and anxiet associated with screening programs. Breast cancer risk assessment has been studied for many years; however, due to reasonably low positive predictive values there are no existing risk models that are universally accepted in routine clinical practice, in particular as related to screening and diagnosis. There is no doubt that a breast cancer risk model with high discriminatory power will enable an increase in efficiency, efficacy, and cost effectiveness of screening paradigms. We propose to develop and test an innovative risk predictor that is based primarily on computed image features representing bilateral mammographic tissue density asymmetry between left and right breasts. As important, we will develop and test this predictor using mammograms acquired prior to any depiction of a highly suspicious abnormality leading to a biopsy and/or a verification of cancer. To achieve our objectives we will assemble a large and diverse image database of full-field digital mammography (FFDM) images with sequentially available images and related clinical information. The database will include three groups of cases, namely (1) positive cases that were verified to have cancer one to six years after the first available negative FFDM examination, (2) screening negative cases that have not been recalled during the period of interest, and (3) recalled and/or biopsied cases due to suspicious mammographic findings, but later proven to be negative or benign. Computed bilateral mammographic tissue asymmetry features will be used to develop the new risk predictor. In addition to evaluating the overall classification performance on the entire database, we will investigate the reproducibility of the predictor's results and the relationship between predictor's classification performance and the time lag between a negative FFDM in question and the first recall due to the actual detection of a highly suspicious finding leading to a biopsy and/or a confirmed cancer. We will also assess the impact, if any, of several other commonly used risk factors (i.e., age, family history, and breast density BIRADS) on predictor's performance. A bootstrapping method will be used to compute predictor's performance levels and 95% confidence intervals. By incorporating this risk predictor with other existing risk models, we will investigate the feasibility of improving discriminatory power in predicting risk of individual women developing breast cancer in near-term (<5 years).

R01CA160205

2012-04-01

2017-03-31