An Improved YOLO V5 Model for Pulmonary Nodule Detection with Synthetic Data Generated by GAN

Pulmonary nodules are an early indication of lung cancer, and their detection is crucial in the diagnosis of lung cancer. Many researches have been conducted and some progress has been made, but it is still a challenge since the boundaries of pulmonary nodules are blurred, small in size and diverse in shape, making nodules difficult to detect. Moreover, medical datasets are often scarce and precious. To alleviate the above problems, we propose an improved YOLO V5 framework for pulmonary nodule detection based on synthetic data from multiple generative adversarial networks (GANs). Specifically, we use Resnet50 as the YOLO V5 backbone, which introduces residual connections in extracting pulmonary nodule features, thus reducing the effect of diminishing returns and avoiding the gradient explosion problem. In addition, a deformable convolutional network is presented to take into account the detection ability of different shapes and sizes of pulmonary nodules. Furthermore, we suggest using DCGAN, StyleGAN and SAGAN separately to generate realistic and diverse training images to complement the dataset, thus solving the problem of lack of lung nodule samples. The experimental results show that 78.2% precision and 68.1% recall were achieved using only Resnet50. After augmenting the dataset with GAN and using deformable convolution, both the detection and recall of pulmonary nodules have been improved, achieving the best result with 87.8% precision and 82.7% recall after augmenting the dataset with SAGAN.