Neural Networks for Pattern Recognition
Neural Networks for Pattern Recognition
Image Processing, Analysis, and Machine Vision
Image Processing, Analysis, and Machine Vision
IEEE Transactions on Image Processing
| Hi-index | 0.00 |
In this study, we present a computerized framework to identify the corresponding image pair of a lesion in CC and MLO views, a prerequisite for combining information from these views to improve the diagnostic ability of both radiologists and CAD systems. A database of 126 mass lesons was used, from which a corresponding dataset with 104 pairs and a non-corresponding dataset with 95 pairs were constructed. For each FFDM image, the mass lesions were firstly automatically segmented via a dual-stage algorithm, in which a RGI-based segmentation and an active contour model are employed sequentially. Then, various features were automatically extracted from the lesion to characterize the spiculation, margin, size, texture and context of the lesion, as well as its distance to nipple. We developed a two-step strategy to select an effective subset of features, and combined it with a BANN to estimate the probability that the two images are of the same physical lesion. ROC analysis was used to evaluate the performance of the individual features and the selected feature subset for the task of distinguishing corresponding and non-corresponding pairs. With leave-one-out evaluation by lesion, the distance feature yielded an AUC of 0.78 and the feature subset, which includes distance, ROI-based energy and ROI-based homogeneity, yielded an AUC of 0.88. The improvement by using multiple features was statistically significant compared to single feature performance (p