Lung nodule detection on digital tomosynthesis images: A preliminary study

In this paper, we address the problem of small pulmonary nodule detection on digital tomosynthesis (DT) scans. We propose efficient, domain-specific filters for the enhancement and classification of bright, rounded structures in three-dimensional volumes. First, 61 DT slices per scan are reconstructed from the DT projections by filtered backprojection (FBP). Next, nodule candidates are searched slice-wise calculating the determinant of the Hessian (DoH). Then a large number of false candidates are removed by a supervised classifier. The features for classification include coordinates, image correlation and overlap with vessels. For the segmentation of the vascular tree, a modification of the Frangi filter is employed. The system is evaluated on simulated DT scans generated from a computed tomography database. A subset of the LIDC/IDRI database of 37 scans was used. 42% of nodules could be detected while producing on average 100 false positives per scan. Sensitivity increased to 77% when restricting the search to nodules marked as visible.