D. O. Cosgrove1, C. J. Doré1, C. Cohen-Bacrie2, J.-P. Henry2; 1London/UK, 2Aix-en-Provence/FR
Out of the 224 cases submitted for the statistical analysis, the 2nd lesions in 2 patients were discarded. Therefore, the analysis of reproducibility was achieved on 222 cases.
The reproducibility of SWE size measurements and the maximum and mean elasticity measurements were very high (IOR> 0.93, 0.84 and 0.88 respectively, maximum = 1).
Thirty more cases were discarded from the logistic regression analysis due to missing data. This 192-lesion sample showed 12 BI-RADS® 2 lesions (all benign), 62 BI-RADS® 3 lesions of which 56 were benign and 6 were malignant, 71 BI-RADS® 4 lesions fo which 42 were benign and 29 were malignant, and 47 BI-RADS® 5 lesions (all malignant).
This sample had 82 malignant lesions, leading to a prevalence of breast cancer in this sample population of 43%.
Using the BIRADS®≥4 test alone, we found a sensitivity of 92,7%, a specificity of 61,8%, a positive predictive value of 64,4% and a negative predictive value of 91,9%. When applied to this classification, the logistic regression model lead to a ROC curve with an area under the curve of 0,773 (Fig 1).
When added to the BIRADS®≥4 scores, the maximum and mean elasticity increased the ROC area from 0.773 to 0.925 and 0.917 respectively (Fig 2 and Fig 3).
Adding further features did not improve the performance of the system. The best three-variable model (BIRADS®≥4 + elasticity shape + maximum elasticity) increased the area under the ROC curve to 0.934 (Fig 4). In this model, sensitivity decreased from 92.7% to 87.8%, but specificity increased from 61.8% to 87.3%. The rate of correctly classified lesions increased from 75 to 87.5%.