Spatially localized sparse representations for breast lesion characterization

Abstract

Rationale: The topic of sparse representation of samples in high dimensional spaces has attracted growing interest during the past decade. In this work, we develop sparse representation-based methods for classification of radiological imaging patterns of breast lesions into benign and malignant states.

Methods: We propose a spatial block decomposition method to address irregularities of the approximation problem and to build an ensemble of classifiers (CL) that we expect to yield more accurate numerical solutions than conventional whole-region of interest (ROI) sparse analyses. We introduce two classification decision strategies based on maximum a posteriori probability (BBMAP-S), or a log likelihood function (BBLL-S).

Results: To evaluate the performance of the proposed approach we used cross-validation techniques on imaging datasets with disease class labels. We utilized the proposed approach for separation of breast lesions into benign and malignant categories in mammograms. The level of difficulty is high in this application and the accuracy may depend on the lesion size. Our results indicate that the proposed integrative sparse analysis addresses the ill-posedness of the approximation problem, producing AUC (area under the receiver operating curve) value of 89.1% for randomized 30-fold cross-validation.

Conclusions: Furthermore, our comparative experiments showed that the BBLL-S decision function may yield more accurate classification than BBMAP-S because BBLL-S accounts for possible estimation bias.

Keywords: Breast lesion characterization; CAD/CADx; Sparse analysis.

Figure 1:

Outline of our spatially localized sparse analysis method.

Figure 2:

Examples of a malignant lesion (left) and a benign lesion (right) from the MIAS dataset.

Figure 3:

ROC curves for 64 × 64 ROI size breast lesion characterization using the proposed block-based ensemble method with BBMAP-S (left), and BBLL-S (right) decision functions with 10-fold (top row), 20-fold (second row) and 30-fold (bottom row) cross-validation.