A 3D residual network-based approach for accurate lung nodule segmentation in CT images

Abstract

Finding cancerous tumors before they spread is very beneficial and might potentially save patients' lives. The availability of reliable and automated lung cancer detection devices is crucial for both cancer diagnosis and radiation treatment planning. Because of the abundance of data, the tumor's size fluctuation, and its location, a CT scan of a lung tumor will show poor contrast. Using deep learning for medical image processing to segment CT images for cancer detection is no easy feat. The malignant lung region shall be effectively separated from the healthy chest area by using an optimization approach with the 3D residual network ResNet50. A dense-feature extraction module takes all of the encoded feature maps and uses them to extract multiscale features. A U-Net model decoder solves the vanishing gradient problem, and a residual network encodes the input lung CT slices into feature maps. Several encoders work in tandem with the suggested design. No matter how severe a lung anomaly is, we have trained a model to extract dense characteristics from it. Even under difficult conditions, the experimental results show that the proposed technique swiftly and correctly produces explicit lung areas without post-processing. The improved segmentation result may also aid in reducing the risk, according to the available data. Evaluation results on the LUNA16 public dataset showed that the provided technique successfully segmented images of lung nodules using accuracy, recall rate, dice coefficient index, and Hausdroff.