A new automated method to evaluate 2D mammographic breast density according to BI-RADS® Atlas Fifth Edition recommendations.

OBJECTIVES: Radiologists' visual assessment of breast mammographic density (BMD) is subject to inter-observer variability. We aimed to develop and validate a new automated software tool mimicking expert radiologists' consensus assessments of 2D BMD, as per BI-RADS V recommendations. METHODS: The software algorithm was developed using a concept of Manhattan distance to compare a patient's mammographic image to reference mammograms with an assigned BMD category. Reference databases were built from a total of 2289 pairs (cranio-caudal and medio-lateral oblique views) of 2D full-field digital mammography (FFDM). Each image was independently assessed for BMD by a consensus of radiologists specialized in breast imaging. A validation set of additional 800 image pairs was evaluated for BMD both by the software and seven blinded radiologists specialized in breast imaging. The median score was used for consensus. Software reproducibility was assessed using FFDM image pairs from 214 patients in the validation set to compare BMD assessment between left and right breasts. RESULTS: The software showed a substantial agreement with the radiologists' consensus (unweighted κ = 0.68, 95% CI 0.64-0.72) when considering the four breast density categories, and an almost perfect agreement (unweighted κ = 0.84, 95% CI 0.80-0.88) when considering clinically significant non-dense (A-B) and dense (C-D) categories. Correlation between left and right breasts was high (rs = 0.87; 95% CI 0.84-0.90). CONCLUSIONS: BMD assessment by the software was strongly correlated to radiologists' consensus assessments of BMD. Its performance should be compared to other methods, and its clinical utility evaluated in a risk assessment model. KEY POINTS: • A new software tool assesses breast density in a standardized way. • The tool mimics radiologists' clinical assessment of breast density. • It may be incorporated in a breast cancer risk assessment model.

A comparison between different prediction models for invasive breast cancer occurrence in the French E3N cohort.

Breast cancer remains a global health concern with a lack of high discriminating prediction models. The k-nearest-neighbor algorithm (kNN) estimates individual risks using an intuitive tool. This study compares the performances of this approach with the Cox and the Gail models for the 5-year breast cancer risk prediction. The study included 64,995 women from the French E3N prospective cohort. The sample was divided into a learning (N = 51,821) series to learn the models using fivefold cross-validation and a validation (N = 13,174) series to evaluate them. The area under the receiver operating characteristic curve (AUC) and the expected over observed number of cases (E/O) ratio were estimated. In the two series, 393 and 78 premenopausal and 537 and 98 postmenopausal breast cancers were diagnosed. The discrimination values of the best combinations of predictors obtained from cross-validation ranged from 0.59 to 0.60. In the validation series, the AUC values in premenopausal and postmenopausal women were 0.583 [0.520; 0.646] and 0.621 [0.563; 0.679] using the kNN and 0.565 [0.500; 0.631] and 0.617 [0.561; 0.673] using the Cox model. The E/O ratios were 1.26 and 1.28 in premenopausal women and 1.44 and 1.40 in postmenopausal women. The applied Gail model provided AUC values of 0.614 [0.554; 0.675] and 0.549 [0.495; 0.604] and E/O ratios of 0.78 and 1.12. This study shows that the prediction performances differed according to menopausal status when using parametric statistical tools. The k-nearest-neighbor approach performed well, and discrimination was improved in postmenopausal women compared with the Gail model.

How to evaluate the calibration of a disease risk prediction tool.

To evaluate the calibration of a disease risk prediction tool, the quantity E/O, i.e. the ratio of the expected to the observed number of events, is usually computed. However, because of censoring, or more precisely because of individuals who drop out before the termination of the study, this quantity is generally unavailable for the complete population study and an alternative estimate has to be computed. In this paper, we present and compare four methods to do this. We show that two of the most commonly used methods generally lead to biased estimates. Our arguments are first based on some theoretic considerations. Then, we perform a simulation study to highlight the magnitude of biases. As a concluding example, we evaluate the calibration of an existing predictive model for breast cancer on the E3N-EPIC cohort.