Efficient atypicality detection in chest radiographs

Expert radiologists are able to quickly detect atypical features in chest radiographs because they have developed a sense of what textures and contours are typical for each anatomic region by viewing a large set of “normal” chest radiographs. Our previous work modeled this type of learning with a transductive One-Nearest-Neighbor (1NN) method that was effective for identifying atypical regions in chest radiographs. However, the need to compute distances between the feature vectors extracted from a test image and a very large archive of feature vectors (previously extracted from corresponding anatomical locations in a large set of “normal” chest radiographs) made the 1NN method very computationally intensive. This paper uses an instance selection mechanism based on an Extended Fuzzy C-Means (EFCM) clustering algorithm to reduce the magnitude of this computation. Our results (based on a large set of real-world chest radiographs obtained from Mayo Clinic) indicate that EFCM can substantially reduce the computational cost of the 1NN method, without a substantial drop in the accuracy of its atypicality estimates.