An A.I. classifier derived from 4D radiomics of dynamic contrast-enhanced breast MRI data: potential to avoid unnecessary breast biopsies

OBJECTIVES: Due to its high sensitivity, DCE MRI of the breast (bMRI) is increasingly used for both screening and assessment purposes. The high number of detected lesions poses a significant logistic challenge in clinical practice. The aim was to evaluate a temporally and spatially resolved (4D) radiomics approach to distinguish benign from malignant enhancing breast lesions and thereby avoid unnecessary biopsies. METHODS: This retrospective study included consecutive patients with MRI-suspicious findings (BI-RADS 4/5). Two blinded readers analyzed DCE images using a commercially available software, automatically extracting BI-RADS curve types and pharmacokinetic enhancement features. After principal component analysis (PCA), a neural network–derived A.I. classifier to discriminate benign from malignant lesions was constructed and tested using a random split simple approach. The rate of avoidable biopsies was evaluated at exploratory cutoffs (C(1), 100%, and C(2), ≥ 95% sensitivity). RESULTS: Four hundred seventy (295 malignant) lesions in 329 female patients (mean age 55.1 years, range 18–85 years) were examined. Eighty-six DCE features were extracted based on automated volumetric lesion analysis. Five independent component features were extracted using PCA. The A.I. classifier achieved a significant (p < .001) accuracy to distinguish benign from malignant lesion within the test sample (AUC: 83.5%; 95% CI: 76.8–89.0%). Applying identified cutoffs on testing data not included in training dataset showed the potential to lower the number of unnecessary biopsies of benign lesions by 14.5% (C(1)) and 36.2% (C(2)). CONCLUSION: The investigated automated 4D radiomics approach resulted in an accurate A.I. classifier able to distinguish between benign and malignant lesions. Its application could have avoided unnecessary biopsies. KEY POINTS: • Principal component analysis of the extracted volumetric and temporally resolved (4D) DCE markers favored pharmacokinetic modeling derived features. • An A.I. classifier based on 86 extracted DCE features achieved a good to excellent diagnostic performance as measured by the area under the ROC curve with 80.6% (training dataset) and 83.5% (testing dataset). • Testing the resulting A.I. classifier showed the potential to lower the number of unnecessary biopsies of benign breast lesions by up to 36.2%, p < .001 at the cost of up to 4.5% (n = 4) false negative low-risk cancers. SUPPLEMENTARY INFORMATION: The online version contains supplementary material available at 10.1007/s00330-021-07787-z.