Automatic lung segmentation in CT images using mask R-CNN for mapping the feature extraction in supervised methods of machine learning using transfer learning

Abstract

According to the World Health Organization, severe lung pathologies bring about 250,000 deaths each year, and by 2030 it will be the third leading cause of death in the world. The usage of (CT) Computed Tomography is a crucial tool to aid medical diagnosis. Several studies, based on the computer vision area, in association with the medical field, provide computational models through machine learning and deep learning. In this study, we created a new feature extractor that works as the Mask R-CNN kernel for lung image segmentation through transfer learning. Our approaches minimize the number of images used by CNN’s training step, thereby also decreasing the number of interactions performed by the network. The model obtained results surpassing the standard results generated by Mask R-CNN, obtaining more than 99% about the metrics of real lung position on CT with our best model Mask + SVM, surpassing methods in the literature reaching 11 seconds for pulmonary segmentation. To present the effectiveness of our approach also in the generalization of models (methods capable of generalizing machine knowledge to other different databases), we carried out experiments also with various databases. The method was able, with only one training based on a single database, to segment CT lung images belonging to another lung database, generating excellent results getting 99% accuracy.