ABSTRACT
In this chapter, two public data sets were used for the development and evaluation of computer-aided diagnosis (CAD) systems of lung nodules, and two types of CAD systems were built using these two data sets; one is the CAD system for estimating lung cancer probability (CADx), and the other is the false-positive reduction system of CAD for distinguishing between nodules and nonnodules (CADe). In these types of CAD, nodule segmentation was not used. The CADx system and the false-positive reduction system were built using the LUNGx data set and the LUNA16 data set, respectively. For the CADx system, a conventional type of CADx (feature extraction and machine learning) and a CADx with a deep convolutional neural network (DCNN) were built. The results of the former type show that its performance might be comparable to that of radiologists for estimating lung cancer probability. Three-dimensional DCNN was utilized for the false-positive reduction system, and it successfully classified nodule candidates between nodules and nonnodules. Results of the current study show that CAD of lung nodules can be built without nodule segmentation.
