Detection of cancer area in lung images with the help of deep learning algorithms

With the acceleration of the process of industrialization, and the increasing number of smokers, the number of deaths due to lung cancer remains high. A large amount of clinical data found that lung cancer appeared in the form of lung nodules in the early stage. CT imaging is one of the non-invasive detection methods to find early lung nodules. The use of computer-assisted CT lung nodule detection system can help doctors in early detection and diagnosis. However, due to the difference in radiation doses and machine parameters, the brightness, contrast, or irradiation angle of the image database of different hospitals are different. The training of the detection system cannot improve the performance of lung nodule classification algorithm by adding data. This causes problems such as low data usage, high label training costs, and easy expiration of classification models. This research provides the Convolutional Neural Network (CNN) and its architectures, such as AlexNet, GoogleNet, Resnet18 and Resnet50. We have included transfer learning by using CNN's pre-trained architectures. These architectures are tested with large data sets which are SPIE AAPM lung image dataset, which is freely available on the web. Due to the lack of training data, data augmentation technique has proposed to increase the number of training data artificially. The Convolution Neural Network architectures' performances were evaluated by some matrixes, which are confusion matrix, recall, precision and f1-score. Based on the above work, Matlab R2019b is used to program the pre-trained networks and obtaining the results based on transfer learning