Machine learning techniques for pulmonary nodule computer-aided diagnosis using CT images: A systematic review

Early detection of pulmonary nodules is critical for the prevention and treatment of lung cancer. Concomitant with recent advancements in computer performance and intelligent algorithms, the efficacy of pulmonary nodule computer-aided diagnosis (CAD) has been continuously improving, and various algorithms have been proposed using different datasets. This study systematically analyzed and compared the performance of machine learning algorithms using the same dataset in the diagnosis of pulmonary nodules through a literature review. The widely used LIDC-IDRI dataset and its subset LUNA16 were used as data objects. The SpringerLink, Science Direct, IEEE Xplore, and PubMed scientific databases were searched, and seventy-five papers were analyzed. Deep-learning-based CAD was found to be superior to conventional machine-learning-based CAD in terms of the number of published studies and algorithm performance. The best performances were as follows: feedforward neural network (FNN) and convolutional neural network (CNN) for detecting pulmonary nodules; region-based CNN (R-CNN) for the segmentation of pulmonary nodules; residual neural network (ResNet) for the classification of nodules and non-nodules; and deep neural network (DNN) for the classification of benign and malignancy. To further extend the application of CAD in clinical practice, the appropriate algorithm type should be used based on the characteristics of the task. The CAD process should be divided into logical stages and the optimal algorithm for each stage should be used to increase the reliability of the process. The CAD performance of numerous algorithms on the same dataset is systematically compared and ideas for future exploration are provided.