Lung nodule semantic segmentation with bi-direction features using U-INET

Abstract

It's difficult to detect lung cancer and determine the severity of the disease without a CT scan of the lungs. The anonymity of nodules, as well as physical characteristics such as curvature and surrounding tissue, suggest that CT lung nodule segmentation has limitations. According to the study, a new, resource-efficient deep learning architecture dubbed U-INET is required. When a doctor orders a computed tomography (CT) scan for cancer diagnosis, precise and efficient lung nodule segmentation is required. Due to the nodules' hidden form, poor visual quality, and context, lung nodule segmentation is a challenging job. The U-INET model architecture is given in this article as a resource-efficient deep learning approach for dealing with the problem. To improve segmentation operations, it also includes the Mish non-linearity functions and mask class weights. Furthermore, the LUNA-16 dataset, which included 1200 lung nodules, was heavily utilized to train and evaluate the proposed model. The U-INET architecture outperforms the current U-INET model by 81.89 times and reaches human expert level accuracy.