(Tuesday, 7/16/2019) 1:45 PM - 3:45 PM
Room: Stars at Night Ballroom 2-3
Purpose: Diffusion-weighted imaging (DWI) and T2-weighted (T2w) MRI are useful clinical adjuncts to T1-weighted dynamic contrast-enhanced (DCE)-MRI in breast cancer diagnostic workup. In this study, we aim to develop a multiparametric radiomic machine learning scheme that utilizes information from these three MRI protocols in the task of distinguishing between benign and malignant breast lesions.
Methods: The database consisted of 392 breast lesions (68 benign and 324 malignant). Lesions were automatically segmented using fuzzy C-means methods, and four categories of radiomic features (geometric, DCE, DWI, and T2w) were extracted. Each category of features was used to train a support vector machine classifier separately with five-fold cross-validation including feature selection. Four approaches to utilizing the multiparametric information were investigated: (i) selecting from the full feature set, (ii) pooling features selected within each category, (iii) averaging and (iv) multiplying the posterior probability of malignancy output from the four classifiers trained separately within each category. Classifier performances were obtained by receiver operating characteristic (ROC) analysis and area under the ROC curve (AUC) served as figure of merit.
Results: Classification of benign and malignant lesions using geometric, DCE, DWI, and T2w features separately yielded AUCs of 0.78±0.03, 0.80±0.03, 0.72±0.03, and 0.76±0.03, respectively. The four multiparametric approaches yielded AUCs of 0.87±0.02, 0.83±0.02, 0.84±0.03, and 0.83±0.03, respectively. Statistically significant differences were shown between the first multiparametric approach and all single-category classifiers, between the second multiparametric approach and the DWI-based classifier, as well as between the last two multiparametric approaches and all single-sequence classifiers except the DCE-based classifier.
Conclusion: Conducting feature selection on the full feature set yielded the highest performance among the multiparametric schemes investigated in this study, while pooling features selected from individual categories produced the lowest performance. The complimentary relationships between the different MRI protocols may be hindered when using separate feature selection schemes.
Funding Support, Disclosures, and Conflict of Interest: Funding sources: NIH grant U01CA195564 in the QIN, the RSNA/AAPM Graduate Fellowship. MLG is a stockholder in R2 Technology/Hologic, a cofounder and equity holder in Quantitative Insights, and a shareholder in QView. MLG receives royalties from Hologic, GE Medical Systems, MEDIAN Technologies, Riverain Medical, Mitsubishi, and Toshiba.
Keywords
Breast, MRI, Classifier Design
Taxonomy
IM/TH- Image Analysis (Single modality or Multi-modality): Classification methods
Contact Email
