Assessing the degree of collinearity among the lesion features of the MRI BI-RADS lexicon

作者：Be***orf Matthias^*; Baltzer Pascal A T; Kaiser Werner A

来源：European Journal of Radiology, 2011, 80(3): E322-E324.

DOI：10.1016/j.ejrad.2010.11.030

• BI-RADS
• Breast MRI
• MR mammography
• Collinearity
• Multicollinearity
• Logistic regression

Purpose: To retrospectively assess collinearity among lesion feature of the MRI BI-RADS lexicon. Collinearity denotes a situation in which two or more (independent) variables are correlated to some degree, thus partly conveying the same information. Collinearity may cause problems in the interpretation of logistic regression models.
Materials and methods: We analysed the BI-RADS features of 351 lesions in 325 consecutive patients. Patients with biopsy proven breast disease or treated with chemotherapy were excluded. All lesion features were dichotomised into "present" or "not present". Correlation matrices were generated for mass and non-mass lesions separately, focus lesions were omitted. The phi coefficient was used as measure for correlation.
Results: There were 253 mass (175 malignant, 78 benign), 66 non-mass (21 malignant, 45 benign) and 32 focus (5 malignant, 27 benign) lesions among the study population. The strongest inter-subgroup correlations among mass lesion features were: slow initial enhancement with persistent kinetics, phi = 0.64 (0.56-0.71), rapid initial enhancement with washout kinetics, phi = 0.52 (0.43-0.61) and rapid initial enhancement with persistent kinetics, phi = -0.43 (-0.53 to -0.32). The strongest inter-subgroup correlation among non-mass lesion features were: rapid initial enhancement with washout kinetics, phi = 0.51 (0.30-0.67), slow initial enhancement with persistent kinetics, phi = 0.43 (0.21-0.61) and rapid initial enhancement with persistent kinetics, phi = -0.41 (-0.18 to -0.60).
Conclusion: There is a noticeable overlap of information, especially between kinetic features and initial enhancement types for both, mass and non-mass lesions. This should be considered when generating logistic regression models with the MRI BI-RADS lesion features.

• 出版日期2011-12