Deep Learning Models for Lung Nodule Segmentation: A Comparative Study

Abstract

Lung nodule detection is clinically crucial but challenging and time-consuming. The development of automated segmentation approaches could be helpful. To assess the capability of deep learning methods for lung diagnosis, this paper compares recent deep learning models and evaluates their performance. We implemented several preprocessing steps, including windowing, thresholding, and resizing, to improve the image quality, adjust the image dimension suitable for the network, and focus on specific areas of interest within an image. The evaluation was conducted on the Lung Image Database Consortium (LIDC) dataset using Dice similarity coefficient (DSC) and Hausdorff distance (HD) metrics with model complexity parameters for a multifaceted comparison of the models. The experiments showed that the highest accuracy among the five chosen models (97.80% DSC and 1.29 HD) was reached by the Connected-UNets model, which also has the highest computational complexity. In this paper, we quantitatively evaluated and compared 5 deep learning models namely Salient Attention UNet, Connected-UNets, DDANet, UTNet, and EdgeNeXt. The evidence-based overview of current deep learning achievements for the clinical community investigated in this study can be useful to the research community in developing a new model and, thus, designing computer-aided detection and diagnosis (CAD) systems.