Improving Automated Lung Segmentation in Ct Images by Adding Anomalies Adjacent to the Pleura

Abstract

Automated lung segmentation in anomalous CT images is of preeminent importance for computer-aided diagnosis systems. State-of-the-art methods tend to either miss consolidated anomalies adjacent to the pleura due to the lack of contrast with the surrounding tissues or include parts of the mediastinum in the segmentation mask. Here we improve a recent and fast segmentation approach based on a sequence of Image Foresting Transforms by adding anomalies adjacent to the pleura with mediastinum exclusion. By that, we expect to improve anomaly segmentation inside the lungs. We present the advantages of our method over the original one and another recent approach based on deep learning using different datasets of anomalous CT images. Our approach can be 1.5 times more accurate than both baselines on average.