Methods for increased sensitivity and scope in automatic segmentation and detection of lung nodules in CT images

Abstract

We propose two methods for lung segmentation and nodule detection in the initial stage. The lung segmentation method is based on a combination of masks, flood fill algorithm, and morphological closing operation. The nodule detection method is based on a multi-level thresholding process combined with various feature extraction techniques. The methods are evaluated with the new Lung Image Database Consortium and Image Database Resource Initiative (LIDC-IDRI) dataset, which increases the number of tested cases, and thus the confidence, reliability and robustness of the results as compared to previously reported research. We aimed at detecting all nodule challenges present in the LIDC-IDRI (nodules<3 mm and nodules with 3–30 mm diametric size). Statistically, we have detected 3058 nodules in the initial stage (including solid, non-or part-solid, juxta-vascular, juxta-plural, and well-circumscribed nodules). To the best of our knowledge and given the results presented in other previous works (for the final stage), our results indicate that we can detect more nodules in the initial stage than ever reported, and that we can handle all types of nodule challenges in LIDC-IDRI. Specifically, we have achieved a sensitivity of 85% on the complete LIDC-IDRI dataset. On the smaller dataset of 60 CT scans with 315 nodules, we detected 301 nodules and achieved a sensitivity of 95.5%.