A Novel Computer-Aided Lung Cancer Detection Method Based on Transfer Learning from GoogLeNet and Median Intensity Projections

In this research, a fast, accurate, and stable system of lung cancer detection based on novel deep learning techniques is proposed. A convolutional neural network (CNN) structure akin to that of GoogLeNet was built using a transfer learning approach. In contrast to previous studies, Median Intensity Projection (MIP) was employed to include multi-view features of three-dimensional computed tomography (CT) scans. The system was evaluated on the LIDC-IDRI public dataset of lung nodule images and 100-fold data augmentation was performed to ensure training efficiency. The trained system produced 81% accuracy, 84% sensitivity, and 78% specificity after 300 epochs, better than other available programs. In addition, a t-based confidence interval for the population mean of the validation accuracies verified that the proposed system would produce consistent results for multiple trials. Subsequently, a controlled variable experiment was performed to elucidate the net effects of two core factors of the system - fine-tuned GoogLeNet and MIPs - on its detection accuracy. Four treatment groups were set by training and testing fine-tuned GoogLeNet and Alexnet on MIPs and common 2D CT scans, respectively. It was noteworthy that MIPs improved the network's accuracy by 12.3%, and GoogLeNet outperformed Alexnet by 2%. Lastly, remote access to the GPU-based system was enabled through a web server, which allows long-distance management of the system and its future transition into a practical tool.