Lung Cancer Detection and Classification using Deep Learning Techniques

Abstract

Better health outcomes have been achieved as a result of improvements in the ability to identify lung cancer. are
utilised frequently in the field of medicine to detect lung tumours at an early stage. Deep learning models such as
UNET, Efficient-net, Resnet, VGG-16, and others have been used in a number of studies to improve the accuracy
with which lung cancer may be detected. This paper offers an approach that combines UNET and Efficient-Net
neural networks for the purpose of lung nodule segmentation and classification. The goal of the algorithm is to
improve the detection performance. In order to make use of the massive volume of CT scan pictures that do not
have any pathological diagnoses attached to them, an approach that is feature-extraction based and semi-supervised
is applied. A feature pyramid network (FPN) with the ResNet-50 model is used for feature extraction, and a neural
network classifier is used for predicting unlabeled nodules. This setup allows for semi-supervised learning to take
place. The skip-connections are the primary innovation that UNET brings to the table. These connections grant the
decoder access to the features that the encoder learnt at different scales, which in turn enables accurate localization
of lung nodules. An effective neural network for image classification is produced by Efficient-Net by employing
scaling on all three dimensions of the network (depth, width, and resolution) in conjunction with a compound
coefficient that scales all of the network's dimensions consistently. This work has been evaluated on the LIDC-IDRI
dataset, which is available to the public, and it performs better than the majority of the existing approaches. Issues
such as a high false-positive rate, small nodules, and a wide range of non-uniform longitudinal data are some of the
problems that the suggested algorithm intends to address. The findings of the experiments indicate that this model
has a better level of accuracy than earlier research, coming in at 91.67% of the time.