Room: Exhibit Hall | Forum 4
Purpose: Radiomic features extracted from dynamic contrast-enhanced magnetic resonance (DCE-MR) images are used to classify breast lesions using machine learning. For characterization of repeatability and robustness of radiomics for potential use in multi-omics discovery, foundational work is needed in determining the most suitable classifiers in various classification tasks.
Methods: Radiomic features extracted from DCE-MR breast images of 267 benign lesions and 905 cancers were used for the following classification tasks: (1) malignant versus benign, (2) invasive ductal carcinoma versus ductal carcinoma in situ, three pairs of tasks for presence or absence of hormone receptors ((3) progesterone, (4) estrogen, and (5) human epidermal receptor 2) and (6) triple negative versus presence of at least one hormone receptor. Lesions were segmented using fuzzy c-means and features automatically extracted using previously reported methods. Ten-fold cross-validation was used with stepwise feature selection within each fold. Linear discriminant analysis (LDA) and support vector machines (SVM) (hyperparameters box constraint and kernel scale optimized) were used as classifiers. The area under the receiver operating characteristic curve (AUC) (proper binormal model) served as figure of merit. Superiority, equivalence, and non-inferiority testing compared classification performance for each task between the two classifiers using an equivalence margin of âˆ†AUC = 0.1.
Results: Difference in AUC (AUCSVM â€“ AUCLDA) [95% CI] for the six classification tasks were -0.010 [-0.047, 0.028], -0.001 [-0.051, 0.049], -0.017 [-0.076, 0.041], -0.021 [-0.088, 0.045], -0.066 [-0.136, 0.003], and -0.065 [-0.134, 0.005], respectively. Classification performance between the two classifiers was equivalent for tasks 1-4. While âˆ†AUC failed to reach statistical significance for tasks 5 and 6, we failed to demonstrate equivalence for these tasks.
Conclusion: LDA and SVM classifiers performed equivalently in several classification tasks for breast lesions using radiomic features extracted from DCE-MR images. Future work will expand the investigation to other classifiers and classification tasks.
Funding Support, Disclosures, and Conflict of Interest: NCI U01 CA195564 NCI R15 CA227948 MLG: stockholder in R2 Technology/Hologic, cofounder/equity holder in Quantitative Insights, receives royalties from Hologic, GE Medical Systems, MEDIAN Technologies, Riverain Medical, Mitsubishi, and Toshiba. KD: royalties from Hologic. AE: Research Consultant, QView Medical, Inc. and Quantitative Insights, Inc. JP: Research Consultant, QView Medical, Inc.