Implementation of Different U-Net Architectures for Segmentation of Lung Cancer CT Images

Abstract

The most precarious cancer in humans is lung cancer. With the problems arising in low accuracy and poor effect of lung nodule segmentation, U-Net-based semantic segmentation approaches are widely used. The paper aims to compare the different types of U-Net models, such as U-Net2D, R2U-Net2D, U-Net++, and Attention U-Net to get the best model out of these. The results from the experiments show that U-Net2D gave the best performance with an accuracy of 99.38%, 74.34% mean IOU, and 0.01 binary cross-entropy loss. Also, it is observed that the training and validation accuracy are approximately the same, thus showing no over-fitting problems, which can aid radiologists in detecting pulmonary lung nodules effectively.