Using deep learning networks for classification of lung cancer nodules in CT images

Purpose: One of the foremost common cancers around the world is Lung Cancer (LC) which evaluationof its incidence very important for more robust planning. Computerized Tomography (CT)is important for the diagnosis of lung nodules in carcinoma. The aim of this article is to createan expert system to help doctors reduce human errors. In each CT scan, a considerable amountof image is generated and sent to the doctor. Studying these images with maximum accuracy ina short time is not possible and therefore problems arise. Small cancerous nodules may not bedetected in the early stages, leading to lower survival rates. Therefore, developing an algorithmfor image analysis is necessary and leads to increased accuracy in diagnosis. The assistive diagnosticsystem marks suspicious areas and doctors can have a faster and more accurate assessmentof the patient’s condition by re-analyzing them. The main problem that this project seeks to solveis human errors caused by current clinical methods for lung cancer detection, which are characterizedby low accuracy and speed. The current project seeks to answer the fundamental questionof whether the analysis of CT scan images from Iranian patient samples using a computer andartificial intelligence methods can reduce human errors or not.Method: The location and size of the cancerous gland are among the practical information providedby CT scan images. Medical images of lung cancer patients will be valued as input to thedata algorithm so that machine learning is completed In this paper, open-source datasets, andmulticenter datasets are used. Three CNN architectures (VGG۱۶, VGG۱۹, and InceptionV۳) weredesigned to detect lung nodules and classified them into two malignant or benign groups based ontheir pathological and laboratory results.Results: The output of this study is automated diagnostic software, which is used by customers ofclinics and hospitals involved in the diagnosis of cancer, and will be used in other centers for lungcancer screening. By developing such algorithms, it is possible to prevent patient deaths due toincorrect and delayed diagnoses and the financial damages that Iranian families suffer at individual,economic, and national levels. In addition to self-sufficiency in issuing practical knowledge, itshould be noted that the accuracy of these three architectures (VGG۱۶, VGG۱۹, and InceptionV۳)was ۹۸.۳%, ۹۹.۶%, and ۹۹.۵%, respectively, and there was no difference in sensitivity and specificitybetween larger and smaller nodes.Conclusion: The model’s credibility was evaluated by manual CT evaluation by physicians, andthe performance of the CNN model was found to be better and more accurate than the manualmethod. The results showed that among the CNN architectures, VGG۱۹ had the best performancewith an accuracy of ۹۹.۶% among the three networks.

کلیدواژه ها:

Deep learning ، Lung Cancer ، early diagnosis ، Computed Tomography