Performance of SURF and SIFT Keypoints for the Automated Differentiation of Abnormality in Chest Radiographs

In this work, automated abnormality detection using keypoint information from Speeded-Up Robust feature (SURF) and Scale Invariant Feature Transform (SIFT) descriptors in chest Radiographic (CR) images is investigated and compared. Computerized image analysis using artificial intelligence is crucial to detect subtle and non-specific alterations of Tuberculosis (TB). For this, the healthy and TB CRs are subjected to lung field segmentation. SURF and SIFT keypoints are extracted from the segmented lung images. Statistical features from keypoints, its scale and orientation are computed. Discrimination of TB from healthy is performed using SVM. Results show that the SURF and SIFT methods are able to extract local keypoint information in CRs. Linear SVM is found to perform better with precision of 88.9% and AUC of 91% in TB detection for combined features. Hence, the application of keypoint techniques is found to have clinical relevance in the automated screening of non-specific TB abnormalities using CRs.

Lung Field Segmentation in Chest X-rays: A Deformation-Tolerant Procedure Based on the Approximation of Rib Cage Seed Points

The delineation of bone structures is a crucial step in Chest X-ray image analysis. In the case of lung field segmentation, the main approach after the localization of bone structures is either their individual analysis or their suppression. We prove that a very fast and approximate identification of bone points that are most probably located inside the lung area can help in the segmentation of the lung fields, without the need for bone structure suppression. We introduce a deformation-tolerant region growing procedure. In a two-step approach, a sparse representation of the rib cage is guided to several support points on the lung border. We studied and dealt with the presence of other bone structures that interfere with the lung field. Our method demonstrated very robust behavior even with highly deformed lung appearances, and it achieved state-of-the-art performance in segmentations for the vast majority of evaluated CXR images. Our region growing approach based on the automatically detected rib cage points achieved an average Dice similarity score of 0.92 on the Montgomery County Chest X-ray dataset. We are confident that bone seed points can robustly mark a high-quality lung area while remaining unaffected by different lung shapes and abnormal structures.