Background:Preoperative non-invasive histologic grading of breast cancer is essential. This study aimed to explore the effectiveness of a machine learning classification method based on Dempster-Shafer (D-S) evidence theory for the histologic grading of breast cancer.Methods:A total of 489 contrast-enhanced magnetic resonance imaging (MRI) slices with breast cancer lesions (including 171 grade Ⅰ, 140 grade Ⅱ, and 178 grade Ⅲ lesions) were used for analysis. All the lesions were segmented by two radiologists in consensus. For each slice, the quantitative pharmacokinetic parameters based on a modified Tofts model and the textural features of the segmented lesion on the image were extracted. Principal component analysis was then used to reduce feature dimensionality and obtain new features from the pharmacokinetic parameters and texture features. The basic confidence assignments of different classifiers were combined using D-S evidence theory based on the accuracy of three classifiers: support vector machine (SVM), Random Forest, and k-nearest neighbor (KNN). The performance of the machine learning techniques was evaluated in terms of accuracy, sensitivity, specificity, and the area under the curve.Results:The three classifiers showed varying accuracy across different categories. The accuracy of using D-S evidence theory in combination with multiple classifiers reached 92.86%, which was higher than that of using SVM (82.76%), Random Forest (78.85%), or KNN (87.82%) individually. The average area under the curve of using the D-S evidence theory combined with multiple classifiers reached 0.896, which was larger than that of using SVM (0.829), Random Forest (0.727), or KNN (0.835) individually.Conclusions:Multiple classifiers can be effectively combined based on D-S evidence theory to improve the prediction of histologic grade in breast cancer.

To determine if pharmacokinetic information derived from dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) could improve the diagnostic value of breast disease. A total of thirty-six female patients (range, 30-70 years, mean 48.1±12.6 years) confirmed BIRADS 3~6 underwent DCE-MRI were retrospectively recruited to this study. Modified two-compartment Tofts model and Brix mode were used to obtain relevant tracer kinetic parameters. Mann-Whitney U-test was used to compare the parameters between the BIRADS 3 and 4, BIRADS 4 and 5, BIRADS 5 and 6, respectively. A P value of less than 0.05 was considered to indicate a significant difference. In evaluating significance between BIRADS 3 and BIRADS 4, the statistical analysis showed that Brix-kep (P=0.014), ABrix (P<0.001), TTP (P=0.022) were independent factors associated with the discrepancy. All kinetic parameters we calculated were independent factors associated with the discrepancy between BIRADS 4 and BIRADS 5. ABrix (AUC=0.915) do have a good discriminative power between BIRADS 3 and 4. Tofts-kep (AUC= 0.952), ABrix (AUC= 0.990) do have a good discriminative power between BIRADS 4 and 5.The numeric values of Brix-kep, Brix-kel, ABrix, Tofts-ktrans, Tofts-kep and Slope were higher in high grades (BIRADS 5 and BIRADS 6) than in low grades (BIRADS 3 and BIRADS 4) tumors. ABrix have a good discriminative power between BIRADS 3 and 4. Tofts-kep, ABrix do have a good discriminative power between BIRADS 4 and 5. No significant difference between BIRADS 5 and 6.