Lung Nodule Segmentation and Classification using U-Net and Efficient-Net

Abstract

The ability to detect lung cancer has led to better health outcomes. Deep learning techniques are widely used in the medical field to detect lung tumors at an early stage. Deep learning models such as U-Net, Efficient-Net, Resnet, VGG-16, etc. have been incorporated in various studies to detect lung cancer accurately. To enhance the detection performance, this work proposes an algorithm that combines U-Net and Efficient-Net neural networks for lung nodule segmentation and classification. A feature-extraction-based semi-supervised method is used to take advantage of the huge amount of CT scan images with no pathological labels. Semi-supervised learning is achieved using a feature pyramid network (FPN) with ResNet-50 model for feature extraction and a neural network classifier for predicting unlabelled nodules. The main innovation of U-Net is the skip-connections, which give the decoder access to the features that the encoder learned at various scales and enable accurate localization of lung nodules. Efficient-Net uses depth, width, and resolution scaling, combined with a compound coefficient that uniformly scales all network dimensions, resulting in an efficient neural network for image classification. This work has been evaluated on the publicly available LIDC-IDRI dataset and outperforms most existing methods. The proposed algorithm aims to address issues such as a high false-positive rate, small nodules, and a wide range of non-uniform longitudinal data. Experiment results show this model has a higher accuracy of 91.67% when compared with previous works.