Radiomic analysis of MR cytometry mappings for breast tumor diagnosis

Abstract

MR cytometry that combines multi-diffusion-time acquisition with quantitative biophysical modeling can provide useful microstructural parameters in tumor imaging. However, current statistical analysis remains limited to incorporating only mean values of intratumoral parameters, while neglecting tumor heterogeneity. This study aims to investigate radiomics added value to MR cytometry for differentiating benign and malignant breast lesions. This prospective study enrolled 221 patients with pathologically confirmed breast lesions from two centers (47 benign vs. 174 malignant). All patients underwent 3T MRI including pulsed gradient spin-echo and oscillating gradient spin-echo sequences. Lesions were manually delineated by radiologists. Time-dependent ADC metrics and MR-cytometry-derived metrics were calculated. Radiomic features were extracted from multiple parameter mappings. Six machine learning models were trained using five-fold cross-validation on the training cohort, and evaluated on both training and external test cohorts. Model performance was evaluated by accuracy, sensitivity, specificity, F1 score, Brier score and the area under the receiver operating characteristic curve (AUC). On the training set, whether based on ADC measurements (mean AUC across six classifiers: 0.868 [0.802–0.933] vs. 0.768 [0.673–0.863]), MR cytometry parameters (0.914 [0.853–0.976] vs. 0.857 [0.782–0.931]), or their combinations (0.935 [0.883–0.986] vs. 0.879 [0.817–0.942]), all radiomics-based models outperformed mean-value-based ones. On the external test set, similar results were obtained: 0.801 [0.699–0.903]) vs. 0.717 [0.598–0.835] for ADC measurements, 0.869 [0.800-0.938] vs. 0.799 [0.710–0.888] for MR cytometry parameters, and 0.901 [0.849–0.952] vs. 0.841 [0.770–0.913] for the combination. LR obtained the highest AUC of 0.943 (0.901–0.986). Compared to conventional ADC measurements, MR cytometry provides additional clinical value and combining them can further improve diagnosis. Radiomics can improve diagnostic performances of machine learning models over the current mean-value-based analysis. MR cytometry, as a non-invasive method that better reflects intratumoral heterogeneity than limited biopsy sampling, has shown promising potential in diagnosis of breast tumors.