Improvement of Mammographic Mass Classification Performance Using an Intelligent Data Fusion Method

Purpose: In order to optimally apply the computer-aided diagnosis (CAD) schemes of mammograms as a "second reader" in the clinical practice, we investigated a new intelligent data fusion method to adaptively combine mammographic mass classification scores rated by individual radiologists and computed by a computer-aided diagnosis (CAD) scheme to improve classification performance. Methods: We assembled a testing dataset that involves 224 regions of interest (ROI) selected from mammograms. Half ROIs depict verified malignant masses and half involve benign masses. A CAD scheme was applied to process these ROIs and classify the masses. We also asked three observers namely a senior, a junior radiologist and a radiology resident to independently read and classify these ROIs. We then applied an alpha integration based data fusion method to adaptively combine classification scores between CAD and each observer as well as multiple observers. A receiver operating characteristic (ROC) method was used to analyze and compare the classification performance changes between each observer, CAD, and data fusion results. Results: The computed areas under ROC curves (AUC) and standard errors are 0.858 +/- 0.025, 0.853 +/- 0,026, 0.776 +/- 0.031 for three observers (from a senior radiologist to a radiology resident), and 0.908 +/- 0.021 for CAD, respectively. Using the alpha integration method, fusion of classification results from multiple observers or between the observer and CAD significantly improved classification performance (p < 0.05). In combining classification results of CAD and each of the three observers, the different optimal fusion weighting factors are generated, which increased AUC values by 12.1%, 10.1% and 20.9% for the three observers, respectively. Conclusions: This study demonstrated the feasibility of applying a new data integration method to identify optimal weighting factors to adaptively fuse the classification scores rated by different observers and computed by CAD, which may help eventually establish a personalized scheme of single-reading plus CAD to assist radiologists in diagnosis of mammograms.