WE‐C‐I‐609‐03: Study of Radial Gradient Features in LDA Classifier for Automated CT Lung Nodule Detection

Purpose: To study the use of radial gradient index features by an LDA classifier for false positive reduction in automated CT lung nodule detection. Method and Materials: Our database contains 38 diagnostic CT scans, with a total of 82 lung nodules. A radial gradient index (RGI)‐based approach is used to reduce false positives detected by our automated method. For each CT section a complementary image (an “RGI map”) is generated in which the pixel intensity is proportional to the RGI computed along a circle of chosen diameter d, centered at that pixel. As the RGI is maximum for a perfect circle, an RGI map enhances the intensity of nodules relative to neighboring anatomic structures. For every candidate we calculate a set of three RGI features, for each of five different values of the RGI diameter. We evaluate the performance of the classifier by introducing in turn RGI features corresponding to a particular diameter, together with an optimal set of 9 non‐RGI features determined previously. The results are compared with the performance of the LDA without RGI features. Finally, we use stepwise LDA in order to identify optimal features. Results: The performance for d = [12, 16, 20] is optimal (the sensitivity increases from 70 % to 79 % at 0.5 false detections/section) while for d ⩾ 24 performance decreases. A stepwise LDA was performed for 10 random partitions of the database in order to evaluate the relative weight of different features for classification. This revealed that 8 out of 15 RGI features were included 9 or more times within the optimal feature set. Conclusion: Inclusion of RGI features results in a substantially improved FROC performance, which is consistent with results of stepwise linear discriminant analysis. Conflict of Interest: SGA is a shareholder in R2 Technology, Inc.